Spaces:
Build error
Build error
Commit
·
ad0d661
1
Parent(s):
a5f83aa
update intro
Browse files- app.py +20 -8
- description.html +21 -0
app.py
CHANGED
|
@@ -1,16 +1,20 @@
|
|
| 1 |
import os
|
| 2 |
-
|
| 3 |
import tensorflow as tf
|
| 4 |
import keras.backend.tensorflow_backend as tb
|
| 5 |
-
tb._SYMBOLIC_SCOPE.value = True
|
| 6 |
import numpy as np
|
| 7 |
import gradio as gr
|
| 8 |
import cv2
|
| 9 |
from PIL import Image
|
| 10 |
from tensorflow.keras.models import load_model
|
| 11 |
|
|
|
|
|
|
|
|
|
|
| 12 |
# Get model weights
|
| 13 |
-
os.system(
|
|
|
|
|
|
|
| 14 |
|
| 15 |
# Load model
|
| 16 |
model = None
|
|
@@ -24,7 +28,7 @@ labels = ["Benign", "Malignant"]
|
|
| 24 |
# Helpers
|
| 25 |
def preprocess_image(img_array):
|
| 26 |
# Normalize to [0,1]
|
| 27 |
-
img_array = img_array.astype(
|
| 28 |
img_array /= 255
|
| 29 |
# Check that images are 2D arrays
|
| 30 |
if len(img_array.shape) > 2:
|
|
@@ -46,16 +50,24 @@ def inference(img):
|
|
| 46 |
labels_probs = {labels[i]: float(preds[i]) for i, _ in enumerate(labels)}
|
| 47 |
return labels_probs
|
| 48 |
|
| 49 |
-
|
|
|
|
| 50 |
description = "This model predicts if the given image has benign or malignant symptoms. To use it, simply upload a skin lesion image, or click one of the examples to load them. Read more at the links below."
|
| 51 |
article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2004.06824' target='_blank'>Melanoma Detection using Adversarial Training and Deep Transfer Learning</a> | <a href='https://github.com/hasibzunair/adversarial-lesions' target='_blank'>Github</a></p>"
|
| 52 |
|
| 53 |
-
|
|
|
|
| 54 |
fn=inference,
|
| 55 |
title=title,
|
| 56 |
-
description
|
| 57 |
article=article,
|
| 58 |
inputs="image",
|
| 59 |
outputs="label",
|
| 60 |
examples=examples,
|
| 61 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import os
|
| 2 |
+
import codecs
|
| 3 |
import tensorflow as tf
|
| 4 |
import keras.backend.tensorflow_backend as tb
|
|
|
|
| 5 |
import numpy as np
|
| 6 |
import gradio as gr
|
| 7 |
import cv2
|
| 8 |
from PIL import Image
|
| 9 |
from tensorflow.keras.models import load_model
|
| 10 |
|
| 11 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
|
| 12 |
+
tb._SYMBOLIC_SCOPE.value = True
|
| 13 |
+
|
| 14 |
# Get model weights
|
| 15 |
+
os.system(
|
| 16 |
+
"wget https://github.com/hasibzunair/adversarial-lesions/releases/latest/download/MelaNet.h5"
|
| 17 |
+
)
|
| 18 |
|
| 19 |
# Load model
|
| 20 |
model = None
|
|
|
|
| 28 |
# Helpers
|
| 29 |
def preprocess_image(img_array):
|
| 30 |
# Normalize to [0,1]
|
| 31 |
+
img_array = img_array.astype("float32")
|
| 32 |
img_array /= 255
|
| 33 |
# Check that images are 2D arrays
|
| 34 |
if len(img_array.shape) > 2:
|
|
|
|
| 50 |
labels_probs = {labels[i]: float(preds[i]) for i, _ in enumerate(labels)}
|
| 51 |
return labels_probs
|
| 52 |
|
| 53 |
+
|
| 54 |
+
title = "Melanoma Detection using Adversarial Training and Deep Transfer Learning"
|
| 55 |
description = "This model predicts if the given image has benign or malignant symptoms. To use it, simply upload a skin lesion image, or click one of the examples to load them. Read more at the links below."
|
| 56 |
article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2004.06824' target='_blank'>Melanoma Detection using Adversarial Training and Deep Transfer Learning</a> | <a href='https://github.com/hasibzunair/adversarial-lesions' target='_blank'>Github</a></p>"
|
| 57 |
|
| 58 |
+
|
| 59 |
+
demo = gr.Interface(
|
| 60 |
fn=inference,
|
| 61 |
title=title,
|
| 62 |
+
description=description,
|
| 63 |
article=article,
|
| 64 |
inputs="image",
|
| 65 |
outputs="label",
|
| 66 |
examples=examples,
|
| 67 |
+
)
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
demo.launch(
|
| 71 |
+
# debug=True,
|
| 72 |
+
# enable_queue=True
|
| 73 |
+
)
|
description.html
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<!DOCTYPE html>
|
| 2 |
+
<html lang="en">
|
| 3 |
+
<head>
|
| 4 |
+
<meta charset="UTF-8">
|
| 5 |
+
<title>Title</title>
|
| 6 |
+
</head>
|
| 7 |
+
<body>
|
| 8 |
+
This is a demo of <a href="https://arxiv.org/abs/2004.06824">Melanoma Detection using Adversarial Training and Deep Transfer Learning</a> (Physics in Medicine and Biology, 2020).</br>
|
| 9 |
+
|
| 10 |
+
We introduce an over-sampling method for learning the inter-class mapping between under-represented
|
| 11 |
+
class samples and over-represented samples in a bid to generate under-represented class samples
|
| 12 |
+
using unpaired image-to-image translation. These synthetic images are then used as additional
|
| 13 |
+
training data in the task of detecting abnormalities in binary classification use-cases.
|
| 14 |
+
Code is publicly available in <a href='https://github.com/hasibzunair/adversarial-lesions' target='_blank'>Github</a>.<br>
|
| 15 |
+
|
| 16 |
+
This method was also effective for COVID-19 detection from chest radiography images which led to
|
| 17 |
+
<a href="https://github.com/hasibzunair/synthetic-covid-cxr-dataset">Synthetic COVID-19 Chest X-ray Dataset for Computer-Aided Diagnosis</a>.
|
| 18 |
+
The synthetic images not only improved performance of various deep learning architectures when used as additional training data
|
| 19 |
+
under heavy imbalance conditions, but also detect the target class (e.g. COVID-19) with high confidence.
|
| 20 |
+
</body>
|
| 21 |
+
</html>
|