text
stringlengths 1
40.9k
|
|---|
T. gondii Pruigniund and RH strains were maintained in vitro as previously described [89] . Soluble toxoplasma antigen (sTAg) was prepared from RH strain tachyzoites as previously described [90] . The Pruigniud strain was used for in vitro tachyzoite infections at a ratio of 3:1. The Me49 strain of T. gondii was maintained in infected Swiss Webster and CBA/CaJ mice. For infection, brains from infected mice were removed placed in 3 ml sterile 16PBS and passed 3-5 times through an 18.5 gauge followed by 20.5 and 22.5 gauge needle. The number of cysts in a 30 ml aliquot was determined microscopically. Brain suspensions were adjusted to 100 cysts/ml and mice were infected each with 20 cysts intraperitoneally. Infection studies of C57Bl/6 and AMCase null mice were conducted at least four times with a minimum of 7 biological controls. C57Bl/6, CBA/CaJ (Jackson, Bar Harbor, ME) and Swiss Webster mice (Charles River, Wilmington, MA) were maintained in a pathogen free environment under IACUC established protocols at the University of California Riverside. AMCase-null mice were generated by targeting exon 5 using loxP/CRE recombination as previously described [85] . The AMCase gene deleted mice were of a mixed background, C57BL/6NTac:129SvEvBrd, and were backcrossed to C57/BL6 for at least 10 generations. These mice were generated and maintained under IACUC protocols established by Pfizer. |
A single cell suspension from spleens was prepared by passing through a nylon 40 mm cell strainer (BD, San Jose, CA). Suspensions were washed with RPMI complete (10% FCS, 1%Pennicilin/Streptomycin, 1% Glutamine, 1% Sodium Pyruvate, 1% nonessential amino acids, 0.1% B-mercaptoethanol) (Life Technologies, Grand Island, NY) and centrifuged for 5 minutes at 1200 RPM at 4uC. Red Blood Cells were lysed using 0.86% ammonium chloride solution, centrifuged and resuspended in RPMI complete. BMNCs were prepared as previously described [89] . |
BMNCs or splenic cells were incubated with various conjugated antibodies against CXCR3, CD3, CD4, CD8, CD11b, IL-10, and CD45, (eBioscience, San Diego, CA) and MMR (Biolegend, San Diego, CA). Cells were analyzed using the BD FACSCanto II flowcytometer (BD Biosciences, San Jose, CA) and FlowJo analysis software v.8.7.3 (Treestar Software, Ashland, OR). Cell populations were determined by gating on CD4+, CD8+, CD45 hi / CD11b+ (macrophages) and CD45 int /CD11b+ (microglia) from live cell gate. |
Total RNA from brain tissue samples was extracted with TRIzol reagent (Life Technologies, Grand Island, NY) according to manufacturer's instructions. DNase1 treatment and first strand cDNA synthesis was performed using cDNA synthesis kit (BioRad, Hercules, CA) according to the manufacturer's instructions. CXCR3, CXCL9, CXCL10, CCL2, CCL5, AMCase, Arg1, and Chit1 specific primers for Real Time PCR were purchased from IDT's primer Quest (http://www.idtdna.com/Scitools/ Applications/Primerquest/). Primer sequences were as follows: |
Immediately following excision, brains were bisected sagitally and flash-frozen in cold isopentane. Frozen brains were then put into standard Tissue-Tek cryomold and filled with Optimal Cutting Temperature (OCT) solution (Tissue-Tek, Torrance, CA) and put on dry ice and subsequently stored at 280uC. Serial sections of 10-20 mm were prepared on a standard Cryostat machine (LEICA/CM1850, Simi Valley, CA). Frozen tissue sections were fixed 75% acetone/25% ethanol then blocked in 10% donkey serum prior to incubation with purified antibodies. Purified primary antibodies for Iba-1 (Wako, Richmond, VA), CXCR3 (Life Technologies, Grand Island, NY), MMR (AbD Serotec, Raleigh, NC) arginase-1, AMCase and stabilin-1 (Santa Cruz Biotechnology, Santa Cruz, CA) as well as biotinylated tomato lectin (Sigma-Aldrich, St. Louis, MO) were incubated with tissue samples for 2 h at RT or overnight at 4uC, and followed with appropriate secondary antibodies conjugated to Alexa 488, Alexa 568, or Alexa 647 at 2 mg/mL (Life Technologies, Grand Island, NY). Samples were mounted in Prolong Gold with DAPI (Life Technologies, Grand Island, NY) for nuclear counterstaining. Images were collected on a Leica SP2 scanning confocol microscope (Leica Optics, Germany), and analyzed using Improvision Volocity 5.0 (Perkin-Elmer, Waltham, MA). Distance of cells from cysts were calculated from confocal images of at least 12 cysts and at least 6 AMCase expressing cells per cyst. |
Parasite burden was measured by amplifying the T. gondii genes B1, SAG1, SAG4, or MAG1 by real-time PCR as previously described [71, 89] . |
In vivo peptide blocking C57BL/6 mice were infected i.p. with 10 4 Pruigniund tachyzoites. At day 21, 23, 25, and 27 post infection the animals were injected i.p. with either 0.5 ml a-CXCL10 (0.5 mg/mL), 0.5 mL a-CXCR3 (polyclonal), or 0.5 ml PBS as previously published [76, 91] . The mice were sacrificed on day 28 p.i. and brains were excised for flow cytometric analysis, and parasite burden as described above. Blocking studies were conducted twice with at least 5 biological replicates. |
Supernatants from infected macrophage cultures were added to a 96 well UV plate at 50 ml per sample in triplicates. Urea reagents A and B were mixed from quantichrom urea assay kit (Bioassay systems, Hayward, CA) and 200 ml of mixture added to each well. Included standard was used starting at 50 mg/ml and diluted 2 fold. Samples were incubated for 30 min at room temperature and plates were read at 520 nm to determine urea concentration. |
Femurs and tibias were obtained from 6-12 week old C57BL/6 mice. After euthanasia, the mice were sprayed with 70% ethanol and the femurs and tibias were dissected using scissors. Muscles connected to the bone were removed using scissors, and the femurs were placed into a 50 mL tube containing sterile DMEM on ice. In a tissue culture hood, the bones were washed in sterile DMEM and then both epiphyses were removed using sterile scissors and forceps. The bone marrow was flushed out with a 10 ml syringe filled with BM20 differentiation media (DMEM supplemented with 10% fetal bovine serum, 20% L929 supernatant, 5% horse serum, 100 U/ml penicillin, 100 mg/ml streptomycin, and 2 mM L-glutamine) (Life Technologies, Grand Island, NY) into a 50 mL sterile tube. The tube was vortexed gently and topped off to 50 mL with fresh BM20. 10 mL of cell suspension was plated out on 100 cm untreated dishes and incubated for 7 days at 37uC, 5% CO 2 with fresh media added at day 4. Cells were then washed, counted and plated at 10 6 cells/mL in BM10 media (DMEM supplemented with 10% fetal bovine serum 10% L929 supernatant, 5% horse serum, 100 U/ml penicillin, 100 mg/ml streptomycin, and 2 mM L-glutamine) (Life Technologies, Grand Island, NY) into a 50 mL sterile tube and allowed to rest for 3 days. Macrophages were stimulated overnight with either recombinant IL-4 (10 ng/ml), LPS (50 ng/ml) or IFN-c (100 U/ml) (all from R&D Systems, Minneapolis, MN), stAg (100 mg/ml) or cystAg (100 mg/ml) in complete DMEM. |
To observe the interaction of macrophages and cysts in vitro, cysts were isolated from the brains of chronically infected mice as described above. 50 cysts were added per well to 96 well plates containing 2610 5 bone marrow derived macrophages. Cysts and cells were viewed using a BD HT Pathway 855 microscope (BD Biosciences, San Jose, CA) in a climate-controlled chamber (37uC, 5% CO 2 ). Nine cysts were identified per condition and photographed every 10 minutes for 14 or 16 hours. Movies were compiled using ImageJ software (NIH, Bethesda, MD) and cyst survival time was determined. |
For statistical analysis of survival data, the log-rank and Gehan-Breslow Wilcoxon test was used and involved over 40 C57Bl/6 and 40 AMCase2/2 mice. Acute (0-14 days) and chronic deaths (.14days) were analyzed individually. For all other data, an unpaired, two-tailed Student's t test, or ANOVA test with a 95% confidence interval was used (Prism; GraphPad Software, Inc., La Jolla, CA). All data are represented as means 6 SEM. Figure S3 AAMØ associated with cyst lysis. Confocal fluorescence microscopy of 20 mm brain slices taken from mice at 4 weeks post infection. Imunohistochemical analysis of alternatively activated macrophage (Iba-1, red) as judged by its expression of stabilin-1 (green), adhering closely to a large round cyst. Polarized, to the site of macrophage 'attachment', bradyzoites are seen escaping in an organized fashion towards or into the AAMØ (arrows). (TIF) Figure S4 Chitinase activity is dependent on the presence of chitin and is independent of IL-4 activation. A) Bone marrow derived macrophages were analyzed for chitinase activity. Macrophages were cultured with whole cysts, cysts treated with trichoderma chitinase or media alone. Data are representative of at least 2 individual experiments with a minimum of n = 3 and are represented as mean 6 SEM. B) qRT-PCR was conducted on BMNC to measure YM-1, YM-2, RELM-a, BRP39, IL-4 and IL4Ra. Data are presented as fold increase over naïve. (TIF) Figure S5 AMCase activity associated with cyst. Confocal fluorescence microscopy of 20 mm brain slices taken from mice at 4 weeks post infection. Immunohistochemical analysis of macrophage (Iba-1, green) and AMCase (red), arrows point to AMCase polarized to the cyst wall. (TIF) Figure S6 AMCase2/2 polarization and infection studies. BMDM from WT and AMCase2/2 mice were polarized to A) M2 and B) M1 phenotype as measured by Urea and Greiss assays respectively. C57Bl/6 (WT) and AMCase2/2 mice were infected with the Me49 strain of T. gondii and sacrificed at C) 3 weeks for cyst counts and D) 5 weeks following infection. RNA was isolated from infected brains, reverse transcribed, and the resulting cDNA was analyzed for SAG1, SAG4, and MAG1 transcript levels using qRT-PCR to measure gene expression from tachyzoites, bradyzoites and cysts, respectively. Results are shown as absolute quantitation of copy number using standard curve. Data are representative of at least 3 individual experiments with a minimum of n = 3 and are represented as mean 6 SEM, ** p,0.01, *** p,0.001. (TIF) Figure S7 BALB/c macrophages lyse cysts more quickly than C57Bl/6 macrophages. BMDM from BALB/c and C57Bl/6 mice were cultured with cysts and imaged using an HT pathway microscope for 16 hours. Images were collected every 10 minutes and cyst survival time was calculated. (TIF) Video S1 3D colocalization of AMCase-secreting macrophages and parasite cysts. Three dimensional z-plane progression of 20 mm confocal image taken from mice at 4 weeks post infection. Green: tomato lectin labeling macrophages; red: AmCase; blue: DAPI labeling nuclei; white: anti-Toxoplasma. (MOV) Video S2 Cysts do not lyse in the absence of macrophages. 14 hour time-lapse movie of cysts cultured in the absence of BMDM. Images were collected every 10 minutes and movies were compiled using ImageJ. (MOV) Video S3 Rupture of T. gondii cysts in the presence of chitinase. 14 hour time-lapse movie of cysts co-cultured with WT BMDM and pretreated with 10 mg/ml trichoderma chitinase. Images were collected every 10 minutes and movies were compiled using ImageJ. |
Video S4 Untreated Cysts cultured with bone marrowderived macrophages. 14 hour time-lapse movie of cysts co-cultured with untreated BMDM. Images were collected every 10 minutes and movies were compiled using ImageJ. |
Video S5 Cysts cultured with BMDM and treated with allosamidin. 14 hour time-lapse movie of cysts co-cultured with WT BMDM and treated with 100 mM allosamidin. Images were collected every 10 minutes and movies were compiled using ImageJ. |
Video S6 Cysts cultured with untreated WT BMDM. 16 hour time-lapse movie of cysts co-cultured with BMDM from WT mice. Images were collected every 10 minutes and movies were compiled using ImageJ. (MOV) |
Video S7 Cysts cultured with untreated AMCase-null BMDM. 16 hour time-lapse movie of cysts co-cultured with BMDM from AMCase-null mice. Images were collected every 10 minutes and movies were compiled using ImageJ. |
Video S8 Cysts cultured with WT BMDM pretreated with LPS/IFN-c. 16 hour time-lapse movie of cysts co-cultured with BMDM from WT mice, pre-treated with LPS/IFN-c. Images were collected every 10 minutes and movies were compiled using ImageJ. |
Video S9 Cysts cultured with WT BMDM pretreated with IL-4. 16 hour time-lapse movie of cysts co-cultured with BMDM from WT mice, pre-treated with IL-4. Images were collected every 10 minutes and movies were compiled using ImageJ. |
Video S10 Me49-RFP cysts cultured with WT BMDM labeled with CellTracker green. 16 hour time-lapse movie of Me49-RFP cysts co-cultured with CellTracker green labeled BMDM from WT mice. Images were collected every 10 minutes and movies were compiled using ImageJ. (MOV) |
Video S11 Me49-RFP cysts cultured with AMCase-null BMDM labeled with CellTracker green. 16 hour time-lapse movie of Me49-RFP cysts co-cultured with CellTracker green labeled BMDM from AMCase-null mice. Images were collected every 10 minutes and movies were compiled using ImageJ. (MOV) Automatic Detection and Quantification of Tree-in-Bud (TIB) Opacities from CT Scans This study presents a novel computer-assisted detection (CAD) system for automatically detecting and precisely quantifying abnormal nodular branching opacities in chest computed tomography (CT), termed tree-in-bud (TIB) opacities by radiology literature. The developed CAD system in this study is based on 1) fast localization of candidate imaging patterns using local scale information of the images, and 2) Möbius invariant feature extraction method based on learned local shape and texture properties of TIB patterns. For fast localization of candidate imaging patterns, we use ball-scale filtering and, based on the observation of the pattern of interest, a suitable scale selection is used to retain only small size patterns. Once candidate abnormality patterns are identified, we extract proposed shape features from regions where at least one candidate pattern occupies. The comparative evaluation of the proposed method with commonly used CAD methods is presented with a dataset of 60 chest CTs (laboratory confirmed 39 viral bronchiolitis human parainfluenza CTs and 21 normal chest CTs). The quantitative results are presented as the area under the receiver operator characteristics curves and a computer score (volume affected by TIB) provided as an output of the CAD system. In addition, a visual grading scheme is applied to the patient data by three well-trained radiologists. Interobserver and observer–computer agreements are obtained by the relevant statistical methods over different lung zones. Experimental results demonstrate that the proposed CAD system can achieve high detection rates with an overall accuracy of 90.96%. Moreover, correlations of observer–observer [Formula: see text] , [Formula: see text] and observer–CAD agreements [Formula: see text] , [Formula: see text] validate the feasibility of the use of the proposed CAD system in detecting and quantifying TIB patterns. I NFECTIOUS lung diseases, such as novel swine-origin H1N1 influenza, tuberculosis (TB), etc., are among the leading causes of disability and death all over the world [1] - [3] , [5] . Computed tomography (CT) examination of the lungs during acute respiratory tract infections has become an important part of patient care, both at diagnosis and monitoring progression or response to therapy. Although CT examination serves as a primary (imaging) diagnostic tool for assessing lung infections, visual analysis of CT images is restricted by low specificity for causal infectious organisms and a limited capacity to assess severity and predict patient outcomes [2] . |
Common CT findings associated with respiratory tract infections include tree-in-bud (TIB) nodularity, ground-glass opacities (GGO), random distribution of pulmonary nodules, linear interstitial/bronchovascular thickening, and consolidations [6] . Although none of these visual patterns are specific for one pathogen, the amount of lung volume exhibiting these features could provide insights into the extent or severity of infection. Among these patterns, TIB opacities, represented by thickened bronchial structures surrounded locally by clusters of 2-3 mm micronodules, are associated with inflammation of the small airways (bronchioles), such as in viral or bacterial bronchiolitis, and the increasing sizes of abnormal regions on CT can suggest the progression of disease [6] . Often considered to have a limited differential diagnosis-M TB infection, infection with nontuberculous mycobacteria, viral infection, cystic fibrosis, this pattern is recognized as a CT appearance of many different entities. Unlike the other imaging patterns such as GGO and consolidations, it is an extremely challenging task to detect and quantify the regions with TIB opacities due to interobserver variations and inconsistent visual scoring methods [2] . Therefore, an accurate method for detecting TIB is a critical in computer-assisted detection (CAD) schemes from chest CT. Although the correct diagnosis for TIB pattern is very important, it is also one of the most difficult tasks for radiologists because the contrast of lesions is often low and the disease patterns are very complex. All these limitations suggest that CAD could make a valuable contribution to the management of respiratory tract infections by assisting in the early recognition of pulmonary parenchymal lesions and providing quantitative measures of disease severity. |
Respiratory tract infections, caused by viruses, bacteria, fungi, and parasites, are a major component of global infectious disease mortality. TIB patterns, in particular, usually represent the disease of the small airways such as infectious-inflammatory bronchiolitis as well as bronchiolar luminal impaction with mucus, pus, cells, or fluid causing normally invisible peripheral airways to become visible on CT [7] . Fig. 1 shows typical TIB patterns in a chest CT ( Fig. 1 shows labeled TIB patterns with blue). As its name implies, this pattern resembles a budding tree in CT due to the branching opacities with adjacent centrilobular nodularity [2] . It is not specific for a single disease entity, but suggests pathology in the peripheral airways, which can be associated with air trapping or subsegmental consolidation in the surrounding alveolar airspaces. Because any organism that infects the small airways can cause a TIB pattern, pulmonary infections are its most common cause. |
TIB is difficult to be detected with conventional CAD systems due to high complexity of their irregular shapes, as well as strong textural similarity of micronodules and thickened airways to other normal and abnormal lung structures. Currently, no reported CAD system is capable of automatically detecting a TIB pattern, therefore, which warrants a need for the development of such a system to improve the diagnostic decision process and quantitative measurement of respiratory tract infections. In this paper, we develop a new CAD system to evaluate respiratory tract infections by automatically detecting and quantifying TIB patterns on CT images. |
The main contributions of this study are twofold. 1) A candidate selection method that locates possible abnormal patterns in the images. This process comes from a learning perspective such that the size, shape, and textural characteristics of TIB patterns are learned a priori. The candidate selection process removes large homogeneous regions from consideration which results in a rapid localization of candidate TIB patterns. The local regions enclosing candidate TIB patterns are then used to extract shape and texture features for automatic detection; 2) another novel aspect in this study is to extract Möbius invariant local shape features (i.e., Willmore energy-based features). Extracted local shape features are combined with statistical texture features to classify lung tissues. In addition, we also investigate the extraction and use of different local shape features as compared to the proposed shape features to facilitate local structure analysis. To the best of our knowledge, this is the first study that uses automatic detection of TIB patterns for a CAD system in infectious lung diseases. Since there is no published work on automatic detection of TIB patterns in the literature, we compare our proposed CAD system on the basis of different feature sets previously shown to be successful in detecting lung diseases in general. Early version of this study appeared in [3] , and can be accessed in [8] . |
This paper is organized as follows. Section II explains the methods of the proposed CAD system. We discuss our proposed and conventional feature extraction methods in Section III. Next, we present the feasibility of the proposed CAD system by evaluating the detection and quantification performances in Section IV followed by a discussion and conclusion in Section V and Section VI, respectively. |
The proposed CAD methodology is illustrated in Fig. 2 . First, lungs are segmented from chest CTs. Second, we use locally adaptive scale-based filtering method to detect candidate TIB patterns. Third, segmented lung is divided into local patches in which we extract Möbius invariant shape features and statistical texture features followed by support vector machine (SVM) classification. We extract features from local patches of the segmented lung only if there are candidate TIB patterns in the patches. The details of the proposed methods are described in the following. |
Prior to detection, segmentation is often the first step in CAD systems. In this study, fuzzy connectedness (FC) image segmentation algorithm is used to achieve successful delineations [9] . In FC framework, as illustrated in Fig. 3 , left and right lungs are "recognized" by user-defined or automatically assigned seeds, which initiate FC segmentation. In this study, one seed per lung volume (i.e., left or right) is automatically set by only considering the locations of small intensity valued voxels inside the body region (see [9] for a detailed description of the use of FC in anatomy segmentation). Fig. 3 (middle and right) shows the resulting segmentation of the chest CT given on the left. Although we use FC algorithm to segment lung regions, there are many well-established lung segmentation methods in the literature [9] - [11] , [29] , and [41] such that they could possibly be used as well to accomplish the delineation step. In that sense, we do not have any strict restriction on the choice of segmentation algorithm prior to detection system as long as it successfully segments the lung regions. |
The size/volume of a region occupied by a typical TIB pattern does not usually exceed a few mm 2 /mm 3 . Together with the fact that TIB pattern has a complex shape with varying intensities over discontinues branches (i.e., buds), TIB patterns have intensity characteristics with high variation toward nearby voxels (see Fig. 1 ). In other words, TIB patterns do not constitute sufficiently large homogeneous regions. Thus, TIB patterns are localized only in the vicinity of small homogeneous regions, and their boundaries have high curvatures due to the nature of its complex shape. In the next section, we use these two observations to extract novel characteristic features to detect TIB patterns. |
Our candidate selection method comes from a learning perspective such that we assign every internal voxel of the lung a membership value reflecting the size (i.e., scale) of the homogeneous region that the voxel belongs to. To do this, we use a locally adaptive scale-based filtering method called ball-scale (or b-scale for short) [9] , [16] , [17] . The b-scale is the simplest form of a locally adaptive scale where the scene is partitioned into several scale levels. Every voxel in each scale is assigned the size of the local structure it belongs. For instance, voxels within the large homogeneous objects have highest scale values, and the voxels nearby the boundary of objects have small-scale values. Voxels on the boundary of objects have smallest scale values. Because of these observations, we conclude that TIB Fig. 4(c) . We describe the computation of b-scale patterns and the details of the candidate selection process in the next section. |
There are several advantages to the local scale-based approach. For instance, boundary-and region-based representations of objects are explicitly contained in the scale-based methods. Based on continuity of homogeneous regions, geometric properties of objects (i.e., size information) can be identified, and this new representation is called scale images, i.e., b-scale, tensor-scale (t-scale), generalized-scale (g-scale) images [16] , [18] , [19] . The b-scale model has been shown to be extremely useful in object recognition [17] , image segmentation [9] , [41] , filtering [16] , inhomogeneity correction [20] , and image registration [20] , [21] . In this study, on the other hand, we show how to use b-scale encoding together with a proper scale selection method for detecting candidate abnormality patterns. The main idea in b-scale encoding is to determine the size of local structures at every voxel as the radius of the largest ball centered at the voxel within which intensities are homogeneous under a prespecified region-homogeneity criterion. |
Although the conventional b-scale encoding method is well established for nD images (n ≥ 2), we use 2-D b-scale encoding method in this study because low-resolution CT data do not allow continuous analysis of TIB patterns through lowresolution imaging direction. In the 2-D digital space (Z 2 , ν), a scene C = (C, f ) is represented by a pair where C is a rectangular array of voxels, ν = (ν 1 , ν 2 ) indicates the size of the voxels, and f is a function that assigns to every voxel an image intensity value. A ball B k (c) of radius k ≥ 0 and with center at |
The fraction of object is denoted by F O k (c) and indicates the fraction of the ball boundary occupied by a region which is sufficiently homogeneous with c. |
and W ψ is a homogeneity function [9] . In all experiments, we use a zero-mean unnormalized Gaussian function for W ψ . The size of the local structure is estimated using appearance information of the gray-level images, i.e., region-homogeneity criterion; b-scale scenes contain rough geometric information. |
A detailed description of W ψ and F O k,ν is presented in [9] . The b-scale algorithm works as follows: the ball radius k is iteratively increased starting from one, and the algorithm checks for F O k,ν (c), the fraction of the object containing c that is contained in the ball. When this fraction falls below a predefined threshold, it is considered that the ball contains an object region different from that to which c belongs [16] . This process is repeated for every voxel within the scene. Voxels are assigned their b-scale values discreetly from 1 to r max . 1 In principle, b-scale partitions the scene into several levels based on the size of local structures from 1 to r max . Computing b-scale values for every voxel leads b-scale scenes as shown in Fig. 4 (a). Note also that locally adaptive scale in regions with fine details or in the vicinity of boundaries is small, while it is large in the interior of large homogeneous objects. |
Developing a successful CAD system for infectious lung diseases requires acquisition of representative features characterizing shape and texture of TIB patterns efficiently. Since TIB is a complex shape pattern consisting of curvilinear structures with nodular structures nearby, we propose to use local shape features (derived from geometry of the local structures) combined with gray-level statistics derived from a given local patch (i.e., local window with a predefined size). |
The shape operator is the second-order invariant (or curvature) which determines the original surface. Since it is usually more convenient to work with scalar quantities rather than vectorial shape quantities, symmetric functions of local Hessian matrices are usually used to extract geometric meaning of the surface/shape of interest. Therefore, curvatures play an important role in the representation and recognition of intrinsic shapes. However, similarity of curvature values may not necessarily be equivalent to intrinsic shape similarities, which causes a degradation in recognition and matching performance. |
To overcome this difficulty, we propose to use Willmore energy functional [22] and several different affine invariant shape features parametrically related to the Willmore energy functional. While local shape features characterize the curvilinear and small nodular structures (via Willmore energy), gray-level features characterize background and foreground intensity variation with objects' pose and size for a given local window. Moreover, for comparison purpose, we use different feature sets previously shown to be successful in detecting lung diseases in general. Fig. 5 enlists all the features that we extracted for the proposed CAD system and for the experimental comparison. Details of extracted features are defined in the following. |
The Willmore energy of surfaces plays an important role in digital geometry, elastic membranes, and image processing [23] . It is closely related to Canham-Helfrich model [24] , where a surface energy is defined as |
where α, β, and γ are some constants, H is the mean curvature vector on Σ (area space), K is the Gaussian curvature on ∂Σ (boundary space), and dA is the induced area metrics on Σ. This model is curvature driven, invariant under the group of Möbius transformations (in particular, under rigid motions and scaling of the surface) and shown to be very useful in energy minimization problems [25] . Invariance of the energy under rigid motions leads to conservation of linear and angular momenta, and invariance under scaling plays a role in setting the size of complex parts of the intrinsic shapes (i.e., corners, wrinkles, folds, etc.). In other words, the position, gray-level characteristics, size, and orientation of the pattern of interest have minimal effect on the extracted features as long as the suitable patch is reserved for the analysis. In order to have simpler and more intuitive representation of the given model, we simply set α = 0 and β = γ = 1, and the equation turns into Willmore energy functional |
where ds is the length metric on ∂Σ. The resultant energy of a surface can be regarded as a function H and K, and captures the deviation of a surface from local sphericity [22] such that a sphere has zero Willmore energy. Note also that the Willmore energy is always nonnegative. Since a homogeneity region that a typical TIB pattern appears is small in size, total curvature (or energy) of that region is high and can be used as a discriminative feature. |
The main motivation in describing intrinsic shapes by Willmore energy is due to its ability to encode surface (i.e., image area in 2-D) with Möbius invariant features (translation, contrast, rotation, and inversion invariant). In addition to Willmore energy features that we adapt from Canham-Helfrich surface model, we have included seven different local shape features, which are parametrically related to Willmore energy formulation, into the proposed CAD system due to their some invariant properties and discriminative powers. Assume κ 1 and κ 2 indicate eigenvalues of the local Hessian matrix H e for any given local patch L , the following shape features are extracted: 1) shape index (SI), 2) Gaussian curvature, 3) mean curvature, 4) elongation, 5) distortion, 6) shear, 7) compactness. |
The SI is a statistical measurement and used to define intrinsic shape of the localized structure within the image [26] , [27] . SI values are encoded as a continuous range of values between −1 and 1, with zero SI indicates saddle-like local structures, +1 and −1 SI values indicate umbilical minima and maxima (i.e., cap and cup, respectively), and midpoints of the two half-intervals (+0.5 and −0.5) indicate concave and convex parabolic or line-like structures (i.e., rut and ridge, respectively). SI can simply be computed through principal curvatures (κ 1 , κ 2 ) as follows: |
where κ 1 ≥ κ 2 . As suggested in [26] , we obtain principal curvatures from the eigenvalues of the local Hessian matrix (H e ) as |
where L xx , L xy = L y x , and L y y are second-order derivatives of local image patch L , and eig() denotes eigenvalue decomposition. We choose to use SI because of its invariance property with respect to rotation, absolute gray value, and translation. |
2) Gaussian Curvature: Gaussian curvature (K) is an intrinsic measure and simply the product of the principal curvatures as K = κ 1 κ 2 for a given point on a surface, equivalent to the determinant of local Hessian matrix H e . Note that K is unchanged even by bending the surface without stretching it, meaning that the Gaussian curvature is independent of the choice of unit normal and it gives three types of classified local shapes: elliptic shape (K > 0), hyperbolic shape (K < 0), parabolic shape (K = 0) with one of the κ is equal to zero, planar shape (K = 0) with both κ are equal to zero. Gaussian curvature is translation and rotation invariant, but not scale invariant. |
3) Mean Curvature: Mean curvature (H) is an extrinsic measure and it describes the curvature as H = (κ 1 + κ 2 )/2. Unlike K, H is defined in the distributional sense. Note that mean curvature measure is the trace of local Hessian matrix H e . Mean curvature can be thought as a negative gradient (as a Laplacian) of the area functional due to its nice variational interpretation over the surface. This does not only give insights into the size of the local shape but also into the total symmetrical deviation from the sphere. Mean curvature is translation and rotation invariant, but not scale invariant. 4) Elongation: Shape elongation is one of the basic shape descriptors and it indicates flatness measure of the local shape [28] . In this paper, we used the ratio of principal curvatures to measure elongation as κ 2 /κ 1 with κ 2 ≤ κ 1 . Elongation measure is invariant with respect to a similarity transformation, and therefore, it is a robust feature that helps to identify curvilinear shapes. Elongation varies from −1 to +1, from hyperbolic to elliptic points. |
Distortion is an algebraic quantity defined as the difference of eigenvalues (i.e., |κ 1 − κ 2 |) of the local Hessian matrix H e . Distortion is a valuable image analysis property revealed by magnitude difference of principal curvatures. Distortion measure captures the deviation of principal curvatures, thus nonplanarity of a region. Together with Gaussian or mean curvature, distortion measure brings further information into encoding of local shape. Distortion measure is translation and rotation invariant, but not scale invariant. 6) Shear: The shear is another algebraic distortion quantity defined as proportional to the normalized distortion: (κ 1 − κ 2 ) 2 /4. The physical information contained in the shear is basically the same as that of the distortion; it is related to distortion with powers of the difference of principal curvatures. Different than distortion, shear descriptor captures higher degree of nonplanarity of a region due to having more robustness against noise. 7) Compactness: Compactness feature measures the similarity between shape of interest and a perfect ellipse, and is defined as 1/(4π √ κ 1 κ 2 ). Note that this ratio is a dimensionless ratio between the area of the shape (1 for a normalized shape) and the area of the best ellipse fitting the shape. Note also that the compactness measure is invariant to affine transformations and parametrically related to Gaussian curvature. Given a single-axial CT slice of left lung, Fig. 6 (b) indicates a thresholded (i.e., selected candidate patterns) b-scale scene encoded from the corresponding gray-level CT slice shown in Fig. 6(a) . Furthermore, Fig. 6 (c) and (d) shows mean and Gaussian curvature maps from which all the other local shape features are extracted, respectively. In addition, Fig. 6 (e) and (f) shows Willmore energy maps using both mean and Gaussian curvature maps as formulated in (4) and shown in Fig. 6(c) and (d) . |
Based on the observation in training step where we analyzed the appearance and shape of TIB patterns, TIB patterns most likely occur in the regions inside the lung with high variability of intensity values over a small number of voxels and with certain size (i.e., a few millimeter in length). These observations (size and high intensity variation) facilitate one practically useful fact in the algorithm that, in the feature extraction process, we only extract features if and only if at least "one" small b-scale pattern exists in the local regions (i.e., blue local regions in Fig. 4) . |
We also explore the use of alternative local shape features as a comparison to Willmore energy-based features. Based on the observations of spatial properties of the selected candidate patterns, it becomes apparent that instead of using conventional high-dimensional feature extraction methods such as Gabor wavelets, steerable wavelets, etc., one may extract much fewer and more reliable statistical features to discriminate the pattern of interest. Motivated from the fact that TIB patterns consist of numerous small (or micro-) nodules nearby the main curvilinear structure and those small structures have varying opacities, the location and distribution of those small structures can be obtained by simple thresholding method which has been popular in estimation for more than two decades [30] . However, since the opacities are varying through different nodular structures, it is challenging to find an optimum threshold value. Therefore, instead of using one single threshold level, we empirically choose n = 10 different threshold levels (λ j ) to obtain local statistics of those structures in a hierarchical way, where λ j = 10j, 1 ≤ j ≤ 10 [31] . This process is named LGS because we extract different statistical measurements in gradient of the images. Note also that we confine ourselves into the local patches where at least one b-scale pattern occupies. |
To obtain shape statistics over local patches, we use gradient fields because boundary information can be used much more effectively in that sense. Fig. 7 shows an example thresholding process over a candidate TIB pattern centered at c (only for four levels are shown for demonstration purpose). After different threshold levels are applied over the local regions of b-scale images, resultant thresholded local patches are used to extract the following features: mean SI values of the local patch for each thresholding level (one feature), and the number of bscale patterns left after thresholding process (one feature). Since we use ten different thresholding levels, we extract 20 features totally. Moreover, for a local region centered at a voxel c of a candidate TIB pattern, we extract one global feature as an SI value of the voxel c, three features as the maximum, minimum, and mean SI values over the local region prior to thresholding. Therefore, a total of 24 features (LGS+SI) are extracted from a typical local patch to be used in CAD system [4] . Although n and λ j are chosen empirically based on the observations of shape and textural characteristics of normal and TIB patterns during the training step, one may propose to use cross validation, control of the global and local false discovery rate, and uncertainty principles to decide those parameters near-optimally [30] . |
Steerable Features: It has been well documented in the literature that decomposition of images by using basis functions localized in spatial position, orientation, and scale (e.g., wavelets) are extremely useful in object recognition and detection [32] , [33] . Since steerable filters are rotation and translation invariant, they accurately represent the underlying image structure [34] . In this study, we use steerable derivative of Gaussian filters to decompose local regions around each candidate pattern into several oriented basis. These basis are used as features in voxelwise classification for TIB identification. We extract steerable features (i.e., directional derivatives) from one scale and six different orientations. |
Gray-Level Co-Occurrence Matrix (GLCM) Features: Spatial statistics based on GLCM [35] are shown to be very useful in discriminating and quantifying patterns pertaining to lung diseases. As texture can give a lot of insights into the classification and characterization problem of poorly defined lesions, regions, and objects, we combine our proposed shape-based invariants with Haralick's popular GLCM-based features [35] . We extract 18 features from each local patch including autocorrelation, contrast, entropy, variance, dissimilarity, homogeneity, cluster shade, energy, maximum probability, sum of squares of variance, sum of averages, sum of variance, sum of entropy, difference of entropy, difference of variance, normalized inverse difference moment, cluster prominence, and mutual information. Readers are encouraged to refer to [35] for further details on these well-established features in machine learning, and [12] - [15] for particular CAD systems in identification of lung abnormalities from CT scans in general. |
Laboratory confirmed (with pathology identification tests) 39 CTs of human parainfluenza (HPIV) infection and 21 normal lung CTs were collected for the experiments. All patients were imaged at our institution using a 64-detector row Philips Brilliance or a 320-detector row Toshiba Aquilion CT scanner. The noncontrasted chest CT studies were performed at end inspiration with 1.0 or 2.0 collimation obtained at 10-or 20-mm intervals from the base of the neck to upper abdomen with a tube voltage of 120 kV and a current of 200-320 mA depending on the subject's weight. Imaging data were constructed to 512 × 512 matrices with slice thickness of 5 mm. The in-plane resolution was affected by patients' size and varied from 0.62 to 0.82 mm. All 60 CT scans (both HPIV and normal) were collected from different subjects (no multiple scans from subjects). |
A well-trained radiologist [with more than nine years experience (DMJ)] carefully examined the complete scan (i.e., 60 CTs) and labeled the lung regions as normal and abnormal (with TIB patterns) (see Fig. 1 ). As many regions as possible showing abnormal lung tissues from 39 HPIV patients were labeled (see Table I for details of the number of regions used in the experiments). Those 39 patients do not include only TIB opacities, but also GGO, nodules, consolidations, and linear thickening such that only TIB regions are labeled in training step. Note also that the control group consisting of 21 subjects with no observed lung abnormalities was constructed and lung tissues pertaining to this group were labeled carefully. |
In the training step, we also explored how the number of b-scale patterns change for normal and diseased subjects. Our observations from detail analysis in candidate selection part showed that only 21-40% of the segmented lung volumes were chosen as candidate TIB patterns. This interval was subject to change based on the severity of the diseases. For patients without having infections (i.e., control group), for instance, the percentage of the candidate regions was smaller than the patients with infections; therefore, an increase in the amount of small-sized b-scale patterns is observed. In any case, local scale could be used as a quantitative measure validating the sensitivity and specificity of the classification rates as we describe it in Section IV-E. |
Occurrences of TIB abnormality and normality of subjects were noted for each CT scan. To analyze existence and severity of abnormality as well as normality of subjects, a visual grading system was adapted from studies examining CT findings in other infections [36] - [38] . Each lung was divided into three zones (for a bilateral total of six) as shown in Fig. 8 . Zone 1 included the apex to the carina. Zone 2 extended from the tracheal carina to the left atrium's junction with inferior pulmonary veins. Zone 3 included the remainder of the lungs below the level of the inferior pulmonary veins atrial junction. A severity score (0 to 5 such that 0 indicates no abnormality) was assigned to each zone based on the percentage of the zone occupied as listed in Fig. 8 (second row). A total score was also extracted by considering all zones during visual grading. Consensus visual scores 2 from participating radiologists [one with more than nine years of experience (DMJ) and one with more than one year of experience (AW)] on a scale of 0-5 over lungs were recorded and compared with computer scores (of the proposed CAD system). Following the same visual scoring scheme, another participating radiologist [with more than seven years of experience (OA)], who was blinded to the consensus visual scores previously obtained, was involved in the visual grading process to provide information on interobserver variability. |
To measure and evaluate the detection capabilities of a CAD system quantitatively, the area under the receiver operator characteristic (ROC) curves is often used [39] . After the proposed CAD system was tested via twofold cross validations with labeled dataset, we presented ROC curves of the system performances. Table I summarizes the performance of the proposed CAD system as compared to other feature sets. The performances are reported as the areas under the ROC curves (A z ). Note that proposed shape features (i.e., Willmore energy and parametrically related local shape features) alone are superior to other methods even though the dimension of the proposed shape feature is only 8. The best performance is obtained when we combine the proposed shape and GLCM features. This is to be expected because spatial statistics are incorporated into the shape features such that texture and shape features are often complementary to each other. On the other hand, compared to the proposed shape features, the LGS and SI features have lower detection rates because they are not affine (and Möbius) invariant and eventually having difficulty in appreciating the large amount of details of TIB patterns. Another reason is that there is no optimal choice of thresholding process and this may yield less remarkable statistical measurements over local patches. However, the LGS and SI features alone perform better than the high-dimensional conventional features similar to the proposed shape features. This result itself suggests the use of local shape features and their adapted extensions in detection of TIB patterns. |
In what follows, we selected the best window size for each feature set and plotted their ROC curves all in Fig. 9 . Superiority of the proposed shape features is clear in all cases. To have a valid comparison, we repeated candidate selection step for all the methods because we observed that the CAD performances of compared conventional feature sets had much lower accuracies if the candidate selection part was not applied (i.e., proposed method's accuracy was decreased to A z = 0.6803, while the best result of all compared methods were decreased to A z < 0.5281). To show whether the proposed method was significantly different than the other methods, we compared the performances through paired t-tests. p-values of the tests indicate that none of the feature set are significantly correlated with the proposed CAD features such that highest and smallest pvalues are reported as 0.0195 (p < 0.05) and 0.0053 (p < 0.01), respectively. |
Visual scoring by radiologists still lies at the heart of diagnostic decisions, and often used as a validation tool. In this section, we explore the correlation between computer score (i.e., CAD score) and visual scores by participating radiologists. Furthermore, we investigate the effectiveness of the proposed method's ability to roughly discriminate normal and diseased patients by only considering the size of the structures pertaining to lung anatomy. |
Based on the visual grading scheme explained in Section IV-C, we compared the consensus reading of two expert observers (AW and DJM) to another expert observer (OA), who was blinded to the consensus scores. We used Pearson product-moment correlation coefficients to determine interobserver agreement over each zone, left, right, and all lung volumes. The reported correlation ratios are shown in Fig. 10(a) . Note that interobserver agreement correlation values for all TIB measurements were high for all zones and the lung. The lowest agreement seen on the zone 1 may be because subtle abnormalities in this zone may have been given greater visual assessment variance among the observers. Nevertheless, an overall correlation coefficient of R 2 = 0.8848 (p < 0.01) indicates an excellent agreement on the existence of TIB patterns. |
We further analyzed the variability of change of scores of expert radiologists for each subject. For this, we constructed Bland-Altman plot [40] where the limits of observer agreements were indicated by bias ± 1.96 std (bias: average difference, std: standard deviation). In Bland-Altman plot, the difference of the performances was plotted against the average of the performances as shown in Fig. 10(b) . It was noted from this figure that the largest disagreement of the scoring between observers To obtain an overall computer score from the proposed CAD system, on the other hand, TIB regions detected by the CAD system were first labeled automatically during the detection process. Then, a computer score was calculated by averaging the volume occupied by the labeled TIB regions over the whole lung volume. Calculated computer score was then normalized to fit the visual grading scheme explained in Section IV-C. Linear regression model was fitted to all subjects' scores both from computer and the consensus scores of the participating radiologists (DMJ and AW) and Pearson product-moment correlation coefficient was computed for this model. A scatter plot of the linear regression model and the computer-observer agreement correlation is shown in Fig. 11 . It is clear from this plot that visual and quantitative assessments correlate well as indicated by the Pearson product-moment correlation of R 2 = 0.824 (p < 0.01). Finally, we illustrate an example of TIB and non-TIB region classification by expert annotation and computer quantification by our proposed method in Fig. 12(a) and (b) , respectively. |
Scale-based analysis: In addition to visual scoring scheme, we also show the effectiveness of the proposed scale-based method on quantification of the disease extent and identification. Scale-based analysis of the regions occupied by TIB patterns is illustrated in Fig. 13 . A CT slice of a patient with HPIV shows fewer large homogeneous regions (green) with respect to a normal control. It also shows a greater number of small homogeneous regions (yellow and red). (from 1 to 10), we recorded the average number of b-scale patterns. As readily seen from both curves, the existence of TIB patterns was indicated through the small number of highly homogeneous regions (i.e., small number of large b-scale patterns) and large number of less homogeneous regions (i.e., large number of small b-scale patterns). This figure validated the qualitative Fig. 13 . The difference between two curves was at statistically significant level (p < 0.01). |
All programs used in this study were developed using gcc 4.5 (Copyrigth (C) 2010 Free Software Foundation) on a Linux platform (Pardus), and all statistical computations were processed in R (Version 2.12.2) and MATLAB (Copyright (C) 2010 Mathworks). All the programs were executed on an Intel (R) Core(TM) i7 CPU 930 at 2.80 GHz with 12 GB RAM workstation. While segmentation of lung regions from CT scans takes only about 10 s, the b-scale encoding algorithm takes a couple of minutes (average 2 min, at most 5 min). The time required to compute b-scale scenes changes from patient to patient due to different number of slices in CT scans. Details of the computational cost analysis for segmentation of lungs, and feature extractions for particular algorithms are enlisted in Table II. A further feature selection method such as a principal component analysis might be used to reduce the dimension of steerable features that we used only for comparison purposes. Note that the proposed features are having a small number of dimension per local patch; there is not necessarily an additional feature selection method needed; hence, it is outside the scope of this paper. |
Briefly, the whole dataset was randomly divided into training and test sets of 30 CT scans (20 HPIV-10 Normal versus 19 HPIV-11 Normal). Parameters of the SVM classifier were learned based on the CT scans pertaining to training set. SVM regression was based on pixel-wise classification [42] . Followed by feature extraction step, the trained SVM classifier was applied to the test set. Note also that we have used twofold cross-validation technique for training and testing; therefore, we changed the role of training and testing dataset in the second fold. We also noticed that there was no significant changes in training and test performances of SVM classifications if twofold cross validation was changed into n-fold cross-validation system with n > 2. In addition, we have used Efron's bootstrap [43] method (i.e., repeating the experiments 100 times based on the actual data) to assess the variability of the estimated classifications derived from SVM regressions, and provide confidence intervals for ROC curves. |
We used radial basis functions as kernel of SVM, and set to epsilon parameter of SVM as 0.1 [42] . Resulting SVM values of pixels are ranging from 0 to 1. This value indicates the likelihood of a local patch belonging to a certain class (TIB or non-TIB); low ratings indicate a non-TIB region, and high ratings indicate a TIB region. Soon after the SVM values were computed for the entire lung, we changed the cutoff values of SVM (0.5 as default) several times to obtain ROC curves. |
In this paper, we studied a very particular, yet important, pattern of lung abnormality observed in chest CTs. Our proposed detection system is tuned to detect TIB regions from non-TIB regions; therefore, a multiclass classifier (with specifically tuned detection filters for each abnormality class) might be needed as an extension of this study to detect as much abnormality as possible in a whole system. Although such a system will bring its unique challenges into the CAD platform, it would be a valuable second opinion tool for radiologists. As a further step, we are currently investigating combining different imaging patterns pertaining to lung abnormalities as well as clinical laboratory information into our CAD system. |
One question arises as to the use of high-resolution CT (HRCT) scans instead of conventional CT scans in detecting TIB patterns, as well as the effect of using HRCT scans in this process. Although HRCT scans appreciate detection of small nodular patterns, they have more noise and lungs might not be fully covered due to large gaps between slices (i.e., 10-30 mm). Furthermore, at our institution and in many other institutions, the protocol for acute pulmonary infection is 5 mm contiguous slice images of the chest without IV contrast, for which we adapted our CAD method. Nevertheless, the method we present is not data dependent and can be used for HRCT scans as well. |
Considering 2-D computation of b-scale scenes, one may doubt if the algorithm can be extended into 3-D. Based on our observations on appearance and location of TIB patterns over the lung regions and experiences on feature extraction in 3-D, as we stated previously, TIB patterns rarely extend in depth direction for more than a few slices due to constraints of lowresolution imaging direction. Therefore, there is no significant classification rate changes in 3-D; however, there is an increase in computational cost. Nevertheless, 3-D b-scale encoding and feature extraction for a similar pattern detection problem or the same problem with high-resolution images (with thinner slice thickness compared to low-resolution CT images) can readily be combined and used with similar accuracies reported in this study. |
Number of large and small b-scale patterns might perhaps be used to identify other type of abnormality patterns such as GGO and consolidations where we expect to have more large b-scale patterns than small b-scale patterns. Therefore, as an extension of this study, we will tune our proposed methodology with different types of abnormalities to generalize the CAD systems for infectious lung diseases in general. |
Our proposed method is capable of detecting and quantifying TIB patterns very accurately as validated by the statistical tests compared to the expert annotations (i.e., ground truth). Therefore, both in detection and quantification steps, the proposed CAD system will highly possibly be helpful for clinicians as a second opinion tool in routine clinical examinations. |
In this study, we have proposed b-scale-based binary classification approach for automatic TIB pattern detection and quantification from chest CTs. The proposed system integrates 1) fast localization of candidate TIB patterns through b-scale filtering and scale selection, and 2) combined shape and textural features to identify TIB patterns. Note that texture-based recognition methods offer a complementary view to shape-based methods; therefore, the integration of spatial information and the proposed shape features achieves high detection rates. Moreover, our proposed local shape features illustrate the usefulness of the invariant properties, Willmore energy in particular, to analyze TIB patterns in chest CT. We have also compared computer scoring of the proposed CAD system with subjective visual grading. A high correlation between objective (CAD) and subjective (visual grading) scores is obtained, which implies highly satisfactory accuracy of the proposed CAD system. Mannose-binding lectin deficiency and acute exacerbations of chronic obstructive pulmonary disease BACKGROUND: Mannose-binding lectin is a collectin involved in host defense against infection. Whether mannose-binding lectin deficiency is associated with acute exacerbations of chronic obstructive pulmonary disease is debated. METHODS: Participants in a study designed to determine if azithromycin taken daily for one year decreased acute exacerbations had serum mannose-binding lectin concentrations measured at the time of enrollment. RESULTS: Samples were obtained from 1037 subjects (91%) in the trial. The prevalence of mannose-binding lectin deficiency ranged from 0.5% to 52.2%, depending on how deficiency was defined. No differences in the prevalence of deficiency were observed with respect to any demographic variable assessed, and no differences were observed in time to first exacerbation, rate of exacerbations, or percentage of subjects requiring hospitalization for exacerbations in those with deficiency versus those without, regardless of how deficiency was defined. CONCLUSION: In a large sample of subjects with chronic obstructive pulmonary disease selected for having an increased risk of experiencing an acute exacerbation of chronic obstructive pulmonary disease, only 1.9% had mannose-binding lectin concentrations below the normal range and we found no association between mannose-binding lectin concentrations and time to first acute exacerbation or frequency of acute exacerbations during one year of prospective follow-up. Mannose-binding lectin (MBL) is a pattern-recognition collectin that is related to surfactant proteins A and D and has two roles in host defense. MBL activates complement via serine proteases, particularly MBL-associated serine protease-2, with which it circulates. 1,2 MBL is also involved in opsonophagocytosis, binding several types of pathogens to phagocytes via its carbohydrate recognition domain, triggering release of a number of proinflammatory cytokines [3] [4] [5] and facilitating clearance of apoptotic cells. 6 MBL is secreted primarily by the liver and circulates in the serum. Low MBL concentrations occur as a result of one of three single nucleotide polymorphisms on exon 1, but the most common cause of deficiency in Caucasians is the LXP haplotype resulting from polymorphisms of the promoter region of the gene, presenting either as a homozygous mutation or in combination with other haplotypes. [7] [8] [9] Several observations suggest that MBL deficiency may compromise host defense in the lungs. MBL binds to carbohydrates on the surface of a number of respiratory pathogens that are associated with acute exacerbations of chronic obstructive pulmonary disease (COPD), including Haemophilus influenzae, 10, 11 Mycoplasma pneumoniae, 12 and influenza. 13 Patients with cystic fibrosis and low MBL concentrations have decreased lung function, an increased prevalence of Burkholderia cepacia infection, and reduced survival relative to patients with cystic fibrosis and normal MBL concentrations. The difference in survival is particularly notable in those with chronic Pseudomonas aeruginosa infection and those with homozygous mutations in the MBL-2 gene. [14] [15] [16] [17] Low MBL concentrations have been associated with an increased risk of respiratory infections in immunocompetent subjects, [18] [19] [20] [21] with an increased frequency of respiratory syncytial virus infections, 22 and with worse outcomes in patients with community-acquired pneumonia and Streptococcus pneumoniae infections. [23] [24] [25] [26] MBL concentrations in bronchoalveolar lavage fluid obtained from a small number of current and former smokers with COPD are lower than those in fluid from healthy controls and tend to be higher in former smokers than in current smokers, 27, 28 but no association has been found between genotypes producing low MBL concentrations and the prevalence of COPD. 29 However, studies assessing the association of low MBL concentrations with acute exacerbations of COPD (AECOPDs) report conflicting results. [30] [31] [32] Accordingly, we prospectively measured MBL concentrations in a large cohort of subjects with COPD who were at increased risk of experiencing acute exacerbations and followed them for one year while tracking the number of acute exacerbations that occurred. Our hypothesis was that subjects with COPD who had an increased risk of experiencing an AECOPD would have more frequent acute exacerbations during one year of follow-up if they were deficient in MBL than if they were not. |
Subjects were men and women enrolled in a multicenter randomized trial designed to determine if azithromycin, taken daily for one year, decreased the frequency of AECOPDs. 33 Eligibility criteria included age $ 40 years, a clinical diagnosis of COPD, and an increased risk of experiencing an AECOPD based on criteria defined by Niewoehner et al. 34 Patients had to be free of AECOPDs for a minimum of 4 weeks prior to enrollment. Exclusion criteria included a diagnosis of asthma or bronchiectasis, among others. 33 Acute exacerbations were defined as "a complex of respiratory symptoms (increased or new onset) of more than one of the following: cough, sputum, wheezing, dyspnea, or chest tightness with a duration of at least three days requiring treatment with antibiotics or systemic steroids". 34 |
Serum was collected at the time subjects were enrolled in the study when they had not experienced an AECOPD for a minimum of 4 weeks and was stored at −80°C until MBL concentrations were assayed by enzyme-linked immunosorbent assay (R & D Systems, Minneapolis, MN). Samples were diluted 1/500 for this assay and assayed in duplicate wells. At this dilution, the range of the standard curve corresponds to concentrations ranging from 78 ng/mL to 5000 ng/mL. When concentrations were extrapolated above 5450 (n = 7) or below 60 (n = 6), the samples were reassayed at either a 1/2500 or a 1/20 dilution and these new values used as MBL serum concentrations. In one case, a sample was still less than the detection limit at 1/20 dilution and this sample was assigned the concentration of less than 3 pg/mL. |
The MBL concentration that def ines MBL def iciency is debated. Some define deficiency as a serum concentration , 500 ng/mL. 24, 31, 35, 36 Others def ine it as #100 ng/mL, and still others define severe deficiency as #50 ng/mL and partial deficiency as .50 ng/mL but ,1000 ng/mL. 9, 10, 21, 25, 31, [37] [38] [39] [40] The normal value reported by the manufacturer of the assay is 1135 ng/mL with a range of 103-3308 ng/mL (R&D Systems). Because of these uncertainties, we defined MBL deficiency in four ways, ie, #50 ng/mL, #100 ng/mL, #500 ng/mL, and .50 but #1000 ng/mL. |
Azithromycin increases expression of the mannose receptor, and uptake of apoptotic cells by human alveolar macrophages, and decreases recovery of apoptotic bronchial epithelial cells. 27 Accordingly, MBL concentrations from patients randomized to receive azithromycin or placebo were analyzed both separately and together. |
A Cox proportional-hazards model analysis was used with time-to-first-exacerbation as the outcome variable and MBL group (ie, below versus above specified limits) as the primary variable of interest. Bootstrap methods were used to compute confidence intervals for median survival times. Rates of AECOPDs were determined by dividing the number of AECOPDs by person-years of follow-up and were analyzed as a function of MBL concentration using a negative binomial model. MBL concentrations are presented as medians and interquartile ranges. P , 0.05 was considered to be statistically significant. The study (ClinicalTrials.gov number NCT00325897) was approved by the institutional review boards at all participating centers. submit your manuscript | www.dovepress.com |
MBL assays were performed at baseline in 1037 (91%) of the 1142 subjects enrolled in the azithromycin trial. Of these, 909 had experienced one or more AECOPDs in the year prior to enrollment. Duplicate measurements of MBL had coefficients of variations ,15% in all cases and ,5% in most. |
The median MBL concentration for all patients was 918 ng/mL (interquartile range 508-1683 ng/mL, inclusive range 0-8194 ng/mL). The median MBL concentration in subjects randomized to receive azithromycin or placebo was 1008 ng/mL (95% confidence interval [CI] 909-1082) and 850 ng/mL (95% CI 776-929), respectively (P = 0.017). |
Patient demographics and clinical characteristics, stratified by MBL concentration, are summarized in Table 1 . The prevalence of MBL deficiency was 0.5%, 1.9%, 24.2%, or 52.3%, when deficiency was defined as #50, #100, #500, or .50 and #1000 ng/mL, respectively (Table 1) . Regardless of the concentration of MBL used to define MBL deficiency, no difference was observed with respect to the prevalence of MBL deficiency by gender, race, or age (with the exception that a greater fraction of women had MBL deficiency defined as .50 ng/mL and #1000 ng/mL than was seen with the other definitions), smoking status, chronic bronchitis, or steroid use, and there was no suggestion that airflow limitation was worse or that GOLD (Global Initiative for Chronic Obstructive Lung Disease) stage was higher in subjects with MBL deficiency ( Table 1 ). The same findings were also observed for the subgroups of subjects receiving azithromycin or placebo (data not shown). |
When analyzing time-to-first AECOPD using a model that included treatment group (ie, azithromycin versus placebo) and log-transformed MBL concentration stratified by clinic, treatment group was significant (P , 0.0001) and log-transformed MBL concentration was not (P = 0.629). The hazard ratio for a one-unit increase in log-transformed MBL concentration was 1.02 (95% CI 0.94-1.12). For the rate per person-year of AECOPDs, a negative-binomial analysis of a model that included treatment group and log-transformed MBL concentration found that treatment group was significant (P = 0.010) but log-transformed MBL concentration was not (P = 0.470). The coefficient for log-transformed MBL concentration in this model was 0.031 (95% CI, −0.053 to +0.115). |
The median time to first AECOPD and the rate of AECOPD per patient-year are shown in Figures 1-3 Table 2 relative to the various definitions of MBL deficiency (only five patients had MBL concentrations # 50 ng/mL, precluding life table analyses for patients in this subgroup). Regardless of the MBL concentration used to define deficiency, no association between the time to first AECOPD, or the rate of AECOPDs and MBL concentration was observed in the population as a whole, or in either treatment subgroup. Nonsignificant trends favoring a longer time to first AECOPD were seen in the subgroup of subjects with MBL concentrations , 100 ng/ mL compared with those with concentrations $ 100 ng/mL ( Figure 1A and C). No difference was observed in MBL concentrations in subjects experiencing no, one, two, or at least three AECOPDs during the course of the study in the population as a whole, or in either treatment group (Table 3) . |
Of the 1037 subjects in the study, 220 (21%) required hospitalization for AECOPDs. The median MBL concentration in these 220 subjects was 1055 (95% CI 861-1213) ng/mL [1091 (958-1450) for those on azithromycin and 891 (816-1276) for those on placebo]. The median MBL concentration in the 817 subjects who were not hospitalized was 904 (833-980, P = 0.17, Table 4 ). |
The important findings of this study are that, in this large sample of subjects with COPD selected for having an increased risk of experiencing an AECOPD within one year, only 1.9% had MBL concentrations below the normal range reported by the manufacturer of the assay, and regardless of the MBL concentration used to define MBL deficiency, we found no association between deficiency and time to first AECOPD, rate of AECOPDs, or need for subjects to be hospitalized for AECOPDs. |
Two studies concluded that MBL deficiency was associated with an increased incidence of AECOPDs and one concluded that it was not. The age, spirometry, and smoking histories of the subjects in these three studies were similar to those we evaluated, with the exception that none of the three studies selected patients who were at increased risk of experiencing acute exacerbations as we did. Yang et al 30 found that 24 of 82 (29%) subjects with COPD had MBL-deficient genotypes. These subjects had lower MBL concentrations than those with the wild-type genotype (107 ng/mL, IQR 30-246, range 21-7675 versus 920 ng/mL, IQR 398-1355, range 21-2256, P , 0.001). MBL concentrations were not presented in a fashion that allowed determination of the prevalence of MBL deficiency based on the definitions used in the literature. Forty of the 82 patients (49%) had one or more admissions for AECOPDs during a two-year follow-up period and 18 of these (45%) had MBL-deficient genotypes. Forty-two patients had no AECOPDs and only six (14%) of these had MBL-deficient genotypes (odds ratio 4.9 95% CI 1.7-14.4, P = 0.0037). We did not determine MBL genotypes but genotypes do not COPD (P = 0.23 and P = 0.10, respectively). No association between MBL concentration and a history of AECOPDs was observed, but no information was provided with respect to how AECOPDs were defined. Eagan et al 31 also noted that subjects with GOLD stage 3 disease had a higher prevalence of MBL deficiency (defined as #100 ng/mL). We found no demographic or COPD severity indicators that were more or less common in subjects with MBL deficiency, regardless of how deficiency was defined ( Table 1) . The strengths of our study include the large sample size, the multicenter design, and the prospective ascertainment of AECOPDs using an event-based (ie, health care utilization) definition. Our study population was enriched by enrolling subjects whom we anticipated would be at increased submit your manuscript | www.dovepress.com risk of experiencing an AECOPD within the one-year follow-up period, based on previous predictors identified by Niewoehner et al. 34 This should have increased the likelihood of finding a higher prevalence of MBL deficiency, regardless of how deficiency was defined, compared with the prevalence in healthy controls. Our study has a number of limitations. First, we only measured MBL concentrations on one occasion. Several groups have demonstrated that MBL is an acute phase reactant. [41] [42] [43] However, in clinically stable patients, MBL concentrations are constant over time, 42, 44 and our patients had to be free of AECOPDs for at least 4 weeks before meeting inclusion criteria. In addition, even during acute phase responses, MBL concentrations in patients with low concentrations only increase by 1.5-4.3-fold and do not reach normal values. 45, 46 Accordingly, we suggest that there is a low likelihood of our data being confounded by spuriously elevated concentrations of MBL resulting in an underestimate of the prevalence of MBL deficiency. |
We did not confirm by genotyping that low MBL concentrations were the result of variant alleles. Approximately 30% and 4%-8% of the normal population have heterozygous or homozygous genetic mutations, respectively, associated with low MBL concentrations. 35 Although MBL concentrations are well correlated with genotypes, (eg, MBL concentrations , 50 ng/mL are 100% sensitive and 83% specific for variant exon-1 polymorphisms 43 ) subjects with wild-type MBL genes may still have low concentrations of MBL, 35 and MBL concentrations may vary as much as 10-fold in patients with identical genotypes for all of the MBL variants described to date. 38 Accordingly, assessing relationships between MBL concentrations and AECOPDs is likely to be a more sensitive approach than genotyping. 21 Similarly, we did not assess MBL function using a complement deposition assay. Bouwman et al 47 and Eisen 48 suggested that while MBL function was a better way of defining MBL deficiency than determining MBL concentrations, assessing MBL concentrations was "most appropriate" for defining MBL deficiency in studies seeking associations with infections. Differences between MBL binding and complement activation may vary depending on the method of assessment, 49 and MBL concentrations from 500 ng/mL to 1000 ng/mL are associated with decreases in function by up to 90%. 9 While we saw no association between MBL deficiency and AECOPDs, MBL deficiency could still be associated with AECOPDs that result from infection with one or more specific pathogens if the frequency with which these specific pathogens affected our subjects was too low to discern the association. However, we think this possibility is unlikely, because the two studies documenting the association between MBL deficiency and AECOPDs found an association with AECOPDs that were not otherwise defined by cause or potential infecting organism. 30, 32 In addition, Lin et al 32 found no difference in the distribution of pathogens in those subjects with AECOPD and without MBL-deficient genotypes. |
Because we only found 20 subjects (1.9%) with MBL concentrations below the lower range of normal reported by the manufacturer of the assay, we cannot exclude the possibility that very low concentrations of MBL have an association with AECOPDs. However, the infrequency of this finding implies that even if this association were found, it would pertain only to a small minority of patients suffering AECOPDs. |
In conclusion, we found a very low prevalence of MBL deficiency in subjects who were at increased risk for experiencing AECOPDs and no association between MBL deficiency and time to first acute exacerbation, frequency of acute exacerbations, or percent of subjects requiring hospitalization for acute exacerbations in subjects with MBL deficiency regardless of how deficiency was defined. Accordingly, our data do not support the idea that MBL might be a therapeutic target to reduce the incidence of AECOPDs. Rather, they imply that, while COPD is an inflammatory disorder with systemic manifestations, the fundamental pathophysiology of COPD differs from conditions in which MBL deficiency seems to be a clear risk factor (ie, childhood pneumonia, rheumatoid arthritis, systemic lupus). The changing phenotype of microglia from homeostasis to disease It has been nearly a century since the early description of microglia by Rio-Hortega; since then many more biological and pathological features of microglia have been recognized. Today, microglia are generally considered to be beneficial to homeostasis at the resting state through their abilities to survey the environment and phagocytose debris. However, when activated microglia assume diverse phenotypes ranging from fully inflamed, which involves the release of many pro-inflammatory cytokines, to alternatively activated, releasing anti-inflammatory cytokines or neurotrophins, the consequences to neurons can range from detrimental to supportive. Due to the different experimental sets and conditions, contradictory results have been obtained regarding the controversial question of whether microglia are “good” or “bad.” While it is well understood that the dual roles of activated microglia depend on specific situations, the underlying mechanisms have remained largely unclear, and the interpretation of certain findings related to diverse microglial phenotypes continues to be problematic. In this review we discuss the functions of microglia in neuronal survival and neurogenesis, the crosstalk between microglia and surrounding cells, and the potential factors that could influence the eventual manifestation of microglia. I. Introduction II. The origin of microglia III.Microlgia the dual natures of neurotoxicity and neuroprotection IV.Crosstalk between microglia and other brain cells 1. Cross talk between microglia and neurons: neurons as regulators of microglial activation 2. Cross talk between astrocytes and microglia: reciprocal influences 3. Microglia-T cell crosstalk:key determinants for the Introduction Microglia are generally considered the immune cells of the central nervous system (CNS) and account for 10% of the total glial cell population in the brain. In a normal physiological environment, they work as sentinel cells by continually screening the brain tissue; they actively participate in pathological processes by changing morphology, expressing various antigens and becoming phagocytic. During the past 20 years, thousands of papers have been published describing both the detrimental and beneficial roles of microglia in various brain disorders, from acute infection or stroke to the long and chronic process of neurodegeneration. Microglia have been firmly established as a key cellular component involved in the eventual outcome of inflammation and eventually contribute to the chronic neurodegeneration; The physiology and signaling of microglia have been comprehensively reviewed by Kettenmann's series papers [1] [2] [3] [4] [5] [6] , however, the regulation of microglial activity is a highly complex system, and the responses of microglia are tailored in a multi-factor dependent manner, and which are the focus we try to review in this paper. |
The precise origin and cell lineage of microglia has been a long time debate. So far two most important hypotheses for microglial origin have been held: "neuroectodermal" or "myeloid-monocytic". Even though the latter has been more widely accepted now, the neuroectodermal hypothesis remains interesting. Skoff [7] detected "multipotential glia cell" with a rat model of optic nerve degeneration and optic nerve development, these cells were demonstrated to originate from neuroectodermal matrix cells, and Kitamura later confirmed this result by describing a continuous morphological transition between glioblasts and ramified microglia in the developing gray matter of hippocampus [8] . The hematopoietic origin of microglia also received a lot attention, the presence of bone marrow Mac-1 positive cells were demonstrated in the brain of embryonic and adult mice, and these cells were proved to be the progenitors for microglial cells [9] , also transplantation of GFP + mice bone marrow cells in GFP-host mice revealed the presence of many GFP + microglia throughout developing and/or inflamed CNS [10, 11] , which strongly suggest the hematopoietic stem cells as one of the origins for replenishment of microglia in the neuropathology. Additionally due to the high similarity in marker expression and phagocytosis behavior between circulating monocytes and microglia, people speculate the monocytic origin of microglia, and a couple of experiments have been performed to show the appearance of labeled monocytes in the developing [12] or inflamed brain [13] . In many cases, the peripheral macrophages are considered to be the orthologue [14, 15] or backup of microglia and infiltrate the brain to supplement microglia, thus to some extent peripheral macrophages mirror the behavior of microglia in the brain and Monocytederived Macrophages (MDMs) from patients have been used as a substitute of microglia in many studies [16] [17] [18] . |
Neuroinflammation has long been considered a mediator of secondary damage following a small injury to the CNS. As the primary immune cells in the brain, microglia are expected to take active roles in the damage process. The presence of activated microglia within injured brain regions and in post-mortem tissue from patients having various neurodegenerative disorders has led to the assumption that all reactive microglia contribute to an adverse and degenerative process. Further studies describe destructive roles for microglia by demonstrating the release of a range of neurotoxins from microglia that includes pro-inflammatory cytokines [19] [20] [21] , nitric oxide [22, 23] and reactive oxygen species [24, 25] ; the inhibition of microglial activation in various experiments results in the attenuation of neurotoxic events and improves neuronal survival. In various neurodegenerative disorders, the over-activation of microglia is considered to be a key causative factor in the process or, at a minimum, to promote the neuropathology. For example, in Alzheimer's disease, microglia activated by amyloid-β(Aβ) protein, the hallmark of the disease, release neurotoxins and potentiate neuronal damage, and this microglial over-activation is an early event that precedes neuropil destruction [26] . The activated microglia cluster around or penetrate the neuritic plaques [27] , supporting a critical role of microglial activation in the pathogenesis and progression of the disease. In Parkinson's disease (PD), an increased number of activated microglia are present in the vicinity of degenerating neurons [28] in the substantia nigra [29] , which is particularly deleterious to dopaminergic neurons due to their glutathione deficiency [30] . A single injection of lipopolysaccharide (LPS) to activate microglia in the substantia nigra region led to a progressive, preferential and irreversible loss of dopaminergic neurons [31] [32] [33] , even though LPS itself did no direct harm to the neurons, indicating that the over-activation of microglia is capable of inducing neuronal death in the absence of other pathological stimulation. All of the evidence described above supports the hypotheses of the neurotoxic features of microglia. |
However, as the sentinel and essential cells of the CNS, it is unlikely that microglia would function to damage neurons in all scenarios. Once stimulated the microglia migrate rapidly to the injury site along the chemokine gradients in vitro [34] and also in response to chemoattractants including ATP and NO released directly or indirectly by the injury [35] to exert effect on the survival of neurons. In fact, some specifically designed experiments have begun to uncover the neuroprotective roles of microglia, and more studies are emerging to show beneficial functions of microglia. Firstly, studies have demonstrated instructive roles for microglia in the developing brain for neuronal differentiation [36, 37] and in the regulation of neuronal apoptosis [38] through the production of neurotrophins [39] . Secondly, in the adult brain, resting microglia, which are characterized by many fine perpendicular processes extending from a few long prolongations, have been regarded as sensor cells for the detection of abnormalities or changes in the brain [40] and help to maintain environmental homeostasis. Lastly but most importantly, activated microglia have also been shown to perform neurotrophic functions following neuronal injury. One compelling study supporting this finding involves the axotomy of peripheral nerves (facial or optic), where a rapid microglial response is exhibited with the efficient clearance of myelin debris that contained inhibitory molecules of axon growth, finally leading to successful axonal regeneration [41] ; the inhibition of this microglial response to facial nerve axotomy impairs neuronal survival [42] . In addition, in neonatal mice administered MPTP, highly activated microglia show neurotrophic potential towards dopamine neurons [43] and after traumatic injury, clear glutamate without evoking inflammatory mediators [44] . The benefits of microglial activation are further demonstrated by the exacerbation of neuropathology in inducible mouse models that are deficient in microglia [45, 46] , the finding of protective microglia in cases of cerebral ischemia [47] and multiple sclerosis [48] and the fact that transplantation of microglia can help to enhance neurite growth and functional recovery after CNS injury [49, 50] . The bunch of factors that can activate microglia and the differential behavior of microglia in various conditions have been listed in Table 1 & 2. The above studies clearly demonstrate that microglia can be neurotrophic in the proper situations; there might be a third possibility that microglia are activated by simply reacting to pathogenic stimulation and takes very limited roles in the neurological disorders, in such case the activation of microglia is solely a result of pathogenic stimulation and work as a bystander that either involved passively during the whole process or even go to apoptosis by some other signals. Thus These activated microglia might have different phenotypes. However, the details of what conditions induce microglia to take beneficial phenotypes remain unknown. Many factors are likely involved in determining the eventual outcome of the manifestation of microglia, including their interaction with neurons or astrocytes in the same environment, age-related dysfunction of microglia, activation timing, and the activation state of the microglia, which we will be discussing below. |
Microglia have been considered to be the first line of defense in the CNS [91] , a hypothesis that has been supported by the finding that microglia actively screen their microenvironment with highly motile processes; thus, the brain is under continual surveillance by microglia. To do this with high efficiency, microglia must be variable, adaptive to their environment and capable of integrating various inputs and responding appropriately [92, 93] . All of these processes require significant interactions with other components within the same environment, including neurons and astrocytes. |
When we talk about whether microglia are neuroprotective or neurotoxic, we only refer to the influence of microglia on neurons. However, many studies indicate that neurons are not merely passive targets of microglia but rather exert control over microglial activities [94] . There are considerable interactions between neurons and microglia. For example, Polazzi hypothesized that activation of microglia as a consequence of neuronal injury is primarily aimed at neuroprotection, with the loss of specific communications between neurons and microglia leading to the neurotoxic behavior of microglia [95] . Accumulating evidence demonstrates that there is significant information exchange between neurons and microglia. Depending on whether they are healthy or injured, neurons send "on" or "off" signals to influence microglial activation. On one hand, the activation of [81, 82] microglia by neuronal injury or degeneration has been widely reported [91, 96] . On the other hand, in the healthy brain, microglial activation is tightly restricted by signaling from neurons. CD200-CD200R has been identified as one of the critical pathways in attenuating microglial activation. CD200 is a member of the immunoglobulin superfamily and is expressed on the neuronal membrane surface, while the CD200 receptor (CD200R) is primarily present in the macrophage lineage, which includes microglia [97] . The disruption of CD200-CD200R interactions results in an accelerated microglial response, whereas intensified CD200-CD200R interactions contribute to an attenuation in neurodegeneration [98] . In mice that have had CD200 selectively removed from neurons, microglia exhibited an activated phenotype and were numerous upon facial nerve transaction; damaged CD200-deficient neurons elicited an accelerated microglial response, which demonstrated a loss of the neuronal inhibitory signal for microglial response [97] . Apart from direct interactions through receptor-ligand combinations, electrical activity and soluble factors released from intact neurons also maintain microglial quiescence. In a neuron-glia co-culture, the blockade of neuronal electrical activity by tetrodotoxin or a glutamate receptor antagonist facilitated microglial activation induced by IFN-γ [99] . Soluble molecules from neurons such as neurotrophins and anti-inflammatory agents downregulate antigen expression on cultured rat microglia [99, 100] . Additionally, released factors from neurons can also influence the survival of microglia. Fukui et al. demonstrated that treatment with conditioned media from mature neurons significantly induced the death of microglial cells independent of LPS, while heated neuron-conditioned media or low-calcium-ion media prevented the death of microglia [101] , indicating that specific factors released from neurons exert detrimental effects on microglia. It has been demonstrated that microglial cells undergo apoptosis following peripheral nerve injury [102] [103] [104] or in cases of experimental autoimmune encephalomyelitis(EAE) [105] Injured neurons induced either neuroprotective or neurotoxic behaviors in microglia depending on the manner of injury [91, [106] [107] [108] [109] , providing strong evidence to support the hypothesis of crosstalk between neurons and microglia. Thus, microglia are not merely surveyors of brain tissue but also receive and actively respond to signals from neurons. |
Although less obvious than the crosstalk with neurons, the interactions between microglia and astrocytes are far from simple and are also crucial for our understanding of how microglia respond to their environment and exert influence on neuronal degeneration or regeneration. Several studies have demonstrated the substantial influence of astrocytes on microglial activation [110] . The induction of microglia by Trimethyltin or Borna disease virus-infected neurons is dependent on the presence of astrocytes [111, 112] . Astrocytes play neuroprotective roles by modulating microglial cell activity and decreasing their cytotoxicity [113, 114] . The expression of IL-12 and the production of inducible nitric oxide synthase (iNOS) in activated microglia have been reported to be suppressed by astrocytes or conditioned media from astrocytes [82, 111, [115] [116] [117] , delineating the signals from astrocytes that affect the activities of microglia. Furthermore, the communication between these two types of cells is two-way; microglia both receive and give signals, as proinflammatory cytokines released from microglia inhibit gap junctions and down-regulate connexin 43 expression in astrocytes [118] [119] [120] , which enhances astrocyte survival. In another study, comparative proteome analysis was performed on astrocytes that were treated with conditioned media from quiescent or activated microglia. Following culture in activated-microglial media, the anti-oxidative enzymes expressed in astrocytes were up-regulated, and these astrocytes were protected against oxidative stress. This result gave insight into the complex intercellular events that take place during neurological disorders [121] . Alzheimer's Disease Internalize and degrade amyloid beta [87] Multiple sclerosis Secrete soluble mediators that trigger neural repair and usually contribute to the creation of an environment conductive for regeneration [48] As in many pathological conditions in the central nervous system such as in neurodegeneration [122] , microglia, activated earlier than astrocytes, promote astrocytic activation through IL-1which is mostly from microglia [123] . On the other hand, activated astrocytes not only facilitate activation of distant microglia via calcium wave [124, 125] , but also inhibit microglial activities [126] . Additionally, it was observed that activated-microglial-conditioned media increased astroglial proliferation [127] , down-regulated the astroglial metabotropic glutamate receptor [128] and induced astroglial brain-derived neurotrophic factor (BDNF) and IL-6 gene expression [129] . Taken together, the importance of microglial activities lies in that they not only exert direct effects on neuronal survival, but they also affect the responses of other supporting cells in the same environment. |
Microglia-T cell crosstalk: key determinants for the trend of immune response |
The entire immune response consists of the cooperation of the innate and adaptive immune systems. In the brain, it has been postulated that the beneficial or destructive outcome of the local microglial (innate) response is determined by a well-controlled dialogue between the innate and the adaptive immune players, which are, in most cases, the microglia and T cells. Activated T cells can cross the bloodbrain barrier and interact with resident microglia in the parenchyma [130] ; these microglia have been characterized as myeloid progenitor cells that can differentiate into macrophage-like or dendritic-like cells [131] and thus work crucially as the principal APCs [85] in the CNS. Monsonego et al. demonstrated that IFN-γ-treated microglia serve as efficient Aβ antigen-presenting cells (APCs) of both Aβ1-40 and Aβ1-42, mediating CD86-dependent proliferation of Aβ-reactive T cells [132] . The activated T cells then exert effects in the injured neural tissues by altering the reactive microglial phenotypes and inducing the astrocytic expression of growth factors or modulating microglia to act as glutamate scavengers [44] to improve neuronal survival [133, 134] . In a model for optic nerve injury, the passive transfer of regulatory CD4 + CD25+ T cells was either destructive or beneficial depending on the genetic background of the mice tested, which determines the differential interaction of T cells with microglia and thus the different T cellmediated microglial phenotypes [133] . Kipnis even observed that both the suppressor and the effector activities of T cells could be mediated through dialogue with microglia in the condition of neurodegneration [135] , The entire scenario of crosstalk between T cells and microglia could be described as the following: microglia are initially activated by pathological stimuli during acute or chronic injury to the brain; if the activation occurs with the proper timing and mode and is well-controlled, the activated microglia will work as APCs [133] to stimulate Treg cells that eventually modulate the microglial activation directly or indirectly and affect the milieu balance between neurotrophism and cytotoxicity [44, 136, 137] . |
Whether microglial activation is neurotrophic or neurotoxic is context-dependent |
After considerable time and research, we have recognized the "double-edged sword" nature of microglial cells. On one hand, significant evidence from in vitro and in vivo studies has associated neuronal injury with microglial activation [138] [139] [140] [141] . This evidence results from an inflammatory phenotype of microglia releasing neurotoxic factors, mediators and reactive oxygen species [138] [139] [140] [141] . On the other hand, several other studies have highlighted the beneficial and important roles of microglia in neuronal regeneration, repair and neurogenesis [142] [143] [144] [145] [146] . These seemingly paradoxical results cannot be directly compared, because they come from different experimental sets that vary in terms of the stimulus, timing of microglial activation and age of animals. Thus, whether microglia have positive or negative effects on neuronal survival is contextdependent. |
There are studies suggesting that senescence in microglia causes them to function abnormally and that the destructive roles of activated microglia in the aged neurodegenerative brain may result from age-associated microglia senescence, causing a failure of the aged microglia to respond correctly to stimuli [147, 148] and eventually promoting neurodegeneration [149] (Figure 1 ). The most prominent and also the initially identified feature of microglial senescence is the morphological alteration described as "dystrophy" [150] . Characteristics of "dystrophic" microglia observed in the aged brain include de-ramification (the loss of finely branched cytoplasmic processes), cytoplasmic beading/ spheroid formation, shortened and twisted cytoplasmic processes, and instances of partial or complete cytoplasmic fragmentation [150] . Such dystrophic microglia were prevalent and extensively distributed in the brain of older human subjects [150, 151] , whereas normally ramified microglial morphology with only rare instances of dystrophic microglia is observed in the young brain [148] . These observations provide initial evidence for the age-associated changes in microglia in the healthy elderly brain. Telomere shortening, a marker of aging, has also been demonstrated in microglia in the aged brain in Flanary's study, who reported that microglial cells in rats exhibit significant telomere shortening and a reduction in telomerase activity during normal aging [152] . More importantly, microglial senescence is also manifested by functional alterations, such as an altered inflammatory profile, increased immunophenotypic expression, and the switch from neuroprotective Luo and Chen Translational Neurodegeneration 2012, 1:9 Page 5 of 13 http://www.translationalneurodegeneration.com/content/1/1/9 in the young brain to neurotoxic in the aged brain upon activation [147] . Also, the timing of microglial proliferation and presentation in the injured aged brain is distinct from that in the young brain. For example, Conde et al. reported that microglial proliferation rates in the aged rat brain were significantly higher than in the young rat brain four days after axotomy of the facial nerve [148] . The distinct pattern of the microglial response to injury in the aged brain has also been recorded in the 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP)-induced model of neurotoxicity [153] , the model of controlled cortical impact (CCI) [154] , cortical stab injury [155] and transient retinal ischemia [156] . Although more attention has been paid to the dysfunction of aged microglia, many critical questions remain unanswered. Some of these questions are: whether the activated state of microglia in the aging brain is concurrent with or secondary to microglial dystrophy; which specific function of microglia is primarily affected by microglial dystrophy, how it is affected and what is the direct consequence of the affected function; and whether the deterioration of a specific microglial function is more related to neurodegeneration than others. Clearly, more research is needed to answer these questions. |
Another important element that critically determines the destructive or neuroprotective role of microglia is the timing of their activation. Because large and very complicated communications pathways exist between immunocompetent cells and cytokines in the CNS, the timing of microglial activation leads to diverse trends and outcomes related to the entire inflammation event. In a model of optic nerve crush injury, Shaked et al. found that an earlier onset of phagocytic activity and antigen presentation by microglia results in a resistance to injury and neurons survived [133] ; the early, moderate, transient and well-controlled activation of local microglia caused them to function as APCs, leading to the communication with Treg cells that subsequently proves to be neuroprotective through the modulation of microglial activation states [133] . In a multiple sclerosis (MS) Luo and Chen Translational Neurodegeneration 2012, 1:9 |
Page 6 of 13 http://www.translationalneurodegeneration.com/content/1/1/9 model of experimental allergic encephalomyelitis (EAE) [157] , the inhibition of microglial activation through tPA knockout (tissue plasminogen activator, an essential element for microglia activation) leads to a delayed onset of the disease but increased severity and delayed recovery from the neurological dysfunction, which suggests that microglial activation is harmful during the onset of the disease but beneficial in the recovery phase [157] . Furthermore, when microglial activation was either stimulated or inhibited at different stages, the disease progression was attenuated or exacerbated accordingly [158] . For example, the inhibition of microglial activation at EAE onset, rather than prior to EAE induction, markedly decreased EAE progression, while the stimulation of microglial activation prior to the onset of EAE promotes lower-level EAE and an earlier recovery from symptoms. Together, these findings suggest different roles for microglial activation during various phases of the disease and that different timing of microglial activation dramatically affects whether microglia will be neuroprotective or deleterious [158] . Similarly, in an oxygen-glucose deprivation model, the time window of microglial neuroprotection has been estimated to up to 48 hour after injury, while the pre-stimulation of microglia with LPS before the injury fails to induce microglial-mediated neuroprotection [86] . It has been proposed that the effects of the early activation of microglia on disease progression could be beneficial through phagocytic activity and antigen presentation, recruitment and interactions with the adaptive immune response and the induction of protective autoimmunity [133] . Furthermore, the balance between protective autoimmunity and autoimmune disease may be determined by the timing and intensity of microglial activation [133] . As the immuno-competent cells in the CNS, microglia are critical determinants of the outcome of injury, and the timing of microglial activation appears to be crucial to the outcome of the injury. Thus, any interference with microglial activation in an attempt to affect the disease course clearly must be temporally-restricted. |
Two distinct phenotypes of macrophages have long been known to play different roles in the inflammatory context. Classically-activated macrophages, characterized by the involvement of T Helper type 1 (Th-1) cytokines such as interferon-γ, promote the release of various proinflammatory cytokines and thus exacerbate the inflammation. Alternatively, activated macrophages predominate in the T Helper type 2 (Th-2) microenvironment and tend to soothe the inflammation. Thus, the behavior of macrophages is dictated by their phenotype, which may eventually affect the beneficial or detrimental roles of macrophages during inflammation. Similarly, research over the past few years has established that microglia do not constitute a single, uniform cell population, but rather comprise a family of cells with diverse phenotypes; some are neuroprotective while others are destructive [92] . So far, three distinct functions have been proposed for microglia. The first is the classical activation state of microglia, which, accompanied by the induction of receptors that participate in the innate immune response [159] , is responsible for the pro-inflammatory milieu, and has been linked to neurotoxic effects in the brain. The second is alternatively activated microglia, which are associated with the production of anti-inflammatory cytokines in the resolution phase of the inflammatory response. Recently, the third activation state of microglia has been identified: it overlaps with and is complementary to the alternative activation and is called acquired deactivation [160, 161] . This is another activation state that promotes immunosuppression and is associated with the anti-inflammatory and functional repair phenotype .Both alternative activation and acquired deactivation down-regulate innate immune responses and have similar gene profiles; the most prominent difference is that acquired deactivation is induced by the exposure of microglia to apoptotic cells or to TGF-β or IL-10, while IL-4 and IL-13 induce alternative activation [160, 161] . It has been observed that multiple activation states of microglia coexist in certain chronic inflammations due to parasitic disease [162] , in which the balance between classical activation and alternative activation/acquired deactivation states is of "benefit" to both host and parasite: the host benefits from reduced self-damage, and the parasite eventually survives within the host. Neurodegenerative disorders are also associated with chronic inflammation and the coexistence of various activation states. For example, in AD, some levels of classical activation may be required to limit the brain levels of Aβ despite the risk of self-damage [163] , while alternative activation of microglia in AD may foster the protection of the surrounding tissue from immune damage even though it may facilitate Aβ deposits. Similar studies [164] [165] [166] have shown that the immune cells in the vicinity of amyloid deposits in AD express mRNA and proteins for pro-inflammatory cytokines, leading to the hypothesis that microglia demonstrate classical activation in AD, while Colton et al. found increased mRNA expression of alternative activation-associated gene profiles in microglia in both the AD brain and an AD mouse model [167] , suggesting the presence of multiple activation states of microglia during neurodegeneration. However, the recognition of heterogeneous phenotypes of microglia only raises more questions: what instructs microglia to acquire a particular phenotype; can any conversion occur between these phenotypes; and is it possible to avoid or at least change the commitment to a destructive phenotype? All of these questions are difficult to answer with our current knowledge of microglia; more extensive work is warranted before we can reach a conclusion. |
Subsets and Splits