convergedmachine/Enhanced-MedMNIST

Cross-Dimensional Evaluation Datasets

Transfer learning in machine learning models, particularly deep learning architectures, requires diverse datasets to ensure robustness and generalizability across tasks and domains. This repository provides comprehensive details on the datasets used for evaluation, categorized into 2D and 3D datasets. These datasets span variations in image dimensions, pixel ranges, label types, and unique labels, facilitating a thorough assessment of fine-tuning capabilities.

Citation:

@online{2411.02441,
Author = {Mehmet Can Yavuz and Yang Yang},
Title = {Cross-D Conv: Cross-Dimensional Transferable Knowledge Base via Fourier Shifting Operation},
Year = {2024},
Eprint = {2411.02441},
Eprinttype = {arXiv},
}

2D Datasets

The 2D datasets span a range of medical imaging modalities and classification tasks. They vary in complexity, from binary classification to multi-class problems, and are standardized to ensure consistent preprocessing. All images have dimensions of (224, 224) and pixel values normalized to the range [0, 255].

Overview of 2D Datasets

Dataset	Modality	Samples	Image Dimensions	Pixel Range	Unique Labels	Label Type
Blood [1]	Microscope	17,092	(224, 224, 3)	0 -- 255	8	Multi-class
Brain [2]	MRI	1,600	(224, 224, 3)	0 -- 255	23	Multi-class
Brain Tumor [3]	MRI	3,064	(224, 224, 3)	0 -- 255	3	Multi-class
Breast Cancer [4]	US	1,875	(224, 224, 3)	0 -- 255	2	Binary
Breast [5]	US	780	(224, 224, 1)	0 -- 255	2	Binary
Derma [6]	Dermatology	10,015	(224, 224, 3)	0 -- 255	7	Multi-class
OrganC [7]	CT	23,582	(224, 224, 1)	0 -- 255	11	Multi-class
OrganS [8]	CT	25,211	(224, 224, 1)	0 -- 255	11	Multi-class
Pneumonia [9]	XR	5,856	(224, 224, 1)	0 -- 255	2	Binary

Insights into 2D Datasets

• Blood: 17,092 microscope images across 8 classes. Excellent for testing models on complex multi-class classification.
• Brain: 1,600 MRI images with 23 labels, providing a challenging multi-class scenario.
• Brain Tumor: 3,064 MRI images in 3 classes, focused on tumor detection and classification.
• Breast Cancer: 1,875 ultrasound images (binary labels), suitable for cancer detection benchmarks.
• Breast: 780 ultrasound images with binary labels, ideal for evaluating performance in small datasets.
• Derma: 10,015 dermatology images across 7 classes, critical for skin lesion classification.
• OrganC & OrganS: 23,582 and 25,211 CT images respectively, focused on organ classification task.
• Pneumonia: 5,856 X-ray images for binary classification of lung infections.

---

3D Datasets

3D datasets provide volumetric data essential for spatial analysis in medical imaging. These datasets test models' capabilities in handling 3D spatial information.

Overview of 3D Datasets

Dataset	Modality	Samples	Image Dimensions	Pixel Range	Unique Labels	Label Type
BraTS21 [10]	MRI	585	(3, 96, 96, 96)	0 -- 22,016	2	Binary
BUSV [11]	US	186	(1, 96, 96, 96)	0 -- 255	2	Binary
Fracture [12]	CT	1,370	(1, 64, 64, 64)	0 -- 255	3	Multi-class
Lung Adenocarcinoma [13]	CT	1,050	(1, 128, 128, 128)	-1,450 -- 3,879	3	Multi-class
Mosmed [14]	CT	200	(1, 128, 128, 64)	0 -- 1	2	Binary
Synapse [15]	Microscope	1,759	(1, 64, 64, 64)	0 -- 255	2	Binary
Vessel [16]	MRA	1,908	(1, 64, 64, 64)	0 -- 255	2	Binary
IXI (Gender) [17]	MRI	561	(2, 160, 192, 224)	0 -- 255	2	Binary

Insights into 3D Datasets

• BraTS21: 585 MRI scans for binary brain tumor classification, testing volumetric analysis.
• BUSV: 186 ultrasound volumes with binary labels, focusing on breast ultrasound imaging.
• Fracture: 1,370 CT volumes in 3 classes, assessing bone fracture detection.
• Lung Adenocarcinoma: 1,050 CT volumes for classifying lung adenocarcinoma subtypes.
• Mosmed: 200 CT volumes for detecting COVID-19-related lung infections.
• Synapse: 1,759 microscope volumes for neural imaging classification.
• Vessel: 1,908 MRA volumes for vessel classification.
• IXI (Gender): 561 MRI volumes labeled by gender, testing demographic classification from brain imaging.

---

Dataset Diversity and Evaluation Suitability

These datasets collectively provide:

• Diverse Modalities: Covering microscopy, CT, MRI, ultrasound, X-ray, and more.
• Wide Complexity Range: From binary classification to multi-class problems.
• Standardized Preprocessing: Uniform image dimensions and pixel scaling.
• Scenarios with Varying Data Size: From small datasets (e.g., BUSV) to large-scale datasets (e.g., OrganS).
• Volumetric Data for 3D Analysis: Testing models' spatial reasoning capabilities.

These datasets are curated to facilitate robust and generalizable machine learning models for real-world medical applications.

---

1. Acevedo et al. (2020)
2. Yavuz et al. (2025)
3. Cheng et al. (2015)
4. Gomez et al. (2024)
5. Al et al. (2020)
6. Tschandl et al. (2018)
7. Xu et al. (2019)
8. Bilic et al. (2023)
9. Kermany et al. (2018)
10. Labella et al. (2023)
11. Lin et al. (2022)
12. Jin et al. (2020)
13. Feng et al. (2020)
14. Morozov et al. (2020)
15. Yang et al. (2020)
16. MedMNIST (v2)
17. IXI Dataset

---

license: mit