Kaynaaf's picture
Update app.py
4142930 verified
import gradio as gr
import matplotlib.pyplot as plt
import keras
from matplotlib.colors import LinearSegmentedColormap
from tensorflow.keras.preprocessing.image import img_to_array
from PIL import Image
import io
import glob
import os
import pickle
from tqdm import tqdm
import numpy as np
with open('model.pkl', 'rb') as f:
model = pickle.load(f)
class_names = ['glioma', 'meningioma', 'notumor', 'pituitary']
def create_custom_colormap():
colors = [(0, 0, 0, 0), (1, 0, 0, 1)]
return LinearSegmentedColormap.from_list('custom', colors)
def create_custom_colormap():
colors = [(0, 0, 0, 0), (1, 0, 0, 1)]
return LinearSegmentedColormap.from_list('custom', colors)
def occlusion_sensitivity(model, img_array, class_index, patch_size=25, stride=5, occlusion_value=0, progress=gr.Progress()):
sensitivity_map = np.zeros((img_array.shape[0], img_array.shape[1]))
original_pred = model.predict(np.expand_dims(img_array, axis=0), verbose=0)[0]
original_prob = original_pred[class_index]
n_steps_h = (img_array.shape[0] - patch_size) // stride + 1
n_steps_w = (img_array.shape[1] - patch_size) // stride + 1
total_steps = n_steps_h * n_steps_w
current_step = 0
for h in range(n_steps_h):
for w in range(n_steps_w):
h_start = h * stride
w_start = w * stride
h_end = min(h_start + patch_size, img_array.shape[0])
w_end = min(w_start + patch_size, img_array.shape[1])
occluded_img = img_array.copy()
occluded_img[h_start:h_end, w_start:w_end, :] = occlusion_value
pred = model.predict(np.expand_dims(occluded_img, axis=0), verbose=0)[0]
prob = pred[class_index]
sensitivity = original_prob - prob
sensitivity_map[h_start:h_end, w_start:w_end] += sensitivity
current_step += 1
progress(current_step / total_steps, desc=f"Analyzing sensitivity: {current_step}/{total_steps}")
return sensitivity_map, original_prob
def create_sensitivity_visualizations(image, sensitivity_map, predicted_class, confidence):
# Original image
fig1, ax1 = plt.subplots(figsize=(6, 6))
ax1.imshow(image, cmap='gray')
ax1.set_title(f'Original Image\n{class_names[predicted_class]} ({confidence:.1%})', fontsize=12, fontweight='bold')
ax1.axis('off')
plt.tight_layout()
buf1 = io.BytesIO()
plt.savefig(buf1, format='png', dpi=150, bbox_inches='tight')
buf1.seek(0)
original_img = Image.open(buf1)
plt.close()
# Sensitivity map
fig2, ax2 = plt.subplots(figsize=(6, 6))
im2 = ax2.imshow(sensitivity_map, cmap='jet')
ax2.set_title('Sensitivity Map', fontsize=12, fontweight='bold')
ax2.axis('off')
plt.colorbar(im2, ax=ax2, fraction=0.046, pad=0.04)
plt.tight_layout()
buf2 = io.BytesIO()
plt.savefig(buf2, format='png', dpi=150, bbox_inches='tight')
buf2.seek(0)
sensitivity_img = Image.open(buf2)
plt.close()
# Overlay
fig3, ax3 = plt.subplots(figsize=(6, 6))
ax3.imshow(image, cmap='gray')
custom_cmap = create_custom_colormap()
masked_sensitivity = np.ma.masked_where(sensitivity_map < 0.15, sensitivity_map)
im3 = ax3.imshow(masked_sensitivity, cmap=custom_cmap, alpha=0.6)
ax3.set_title('Overlay (Red = High Importance)', fontsize=12, fontweight='bold')
ax3.axis('off')
plt.colorbar(im3, ax=ax3, fraction=0.046, pad=0.04)
plt.tight_layout()
buf3 = io.BytesIO()
plt.savefig(buf3, format='png', dpi=150, bbox_inches='tight')
buf3.seek(0)
overlay_img = Image.open(buf3)
plt.close()
return original_img, sensitivity_img, overlay_img
def predict_and_visualize(image, patch_size=25, stride=5, progress=gr.Progress()):
if image is None:
return None, None, None, "⚠️ Please select or upload an image"
if not isinstance(image, Image.Image):
image = Image.fromarray(image)
image = image.convert('L')
image = image.resize((256, 256))
img_array = img_to_array(image) / 255.0
# Initial prediction
predictions = model.predict(np.expand_dims(img_array, axis=0), verbose=0)[0]
predicted_class = np.argmax(predictions)
confidence = predictions[predicted_class]
pred_text = f"## 🎯 Prediction Results\n\n"
pred_text += f"### **Predicted Class:** {class_names[predicted_class].upper()}\n"
pred_text += f"### **Confidence:** {confidence:.2%}\n\n"
pred_text += "---\n\n"
pred_text += "### πŸ“Š Class Probabilities:\n\n"
# Sort predictions by probability
sorted_indices = np.argsort(predictions)[::-1]
for i in sorted_indices:
bar = "β–ˆ" * int(predictions[i] * 20)
pred_text += f"**{class_names[i].capitalize()}:** {predictions[i]:.2%} {bar}\n\n"
# Compute sensitivity map with progress tracking
sensitivity_map, _ = occlusion_sensitivity(
model, img_array, predicted_class,
patch_size=patch_size, stride=stride,
progress=progress
)
# Normalize sensitivity map
eps = 1e-9
sensitivity_map = (sensitivity_map - np.min(sensitivity_map)) / (np.max(sensitivity_map) - np.min(sensitivity_map) + eps)
# Create visualizations
original_img, sensitivity_img, overlay_img = create_sensitivity_visualizations(
image, sensitivity_map, predicted_class, confidence
)
progress(1.0, desc="Complete!")
return original_img, sensitivity_img, overlay_img, pred_text
def get_example_images(examples_folder='examples/'):
"""Get list of example images from the Examples folder"""
if not os.path.exists(examples_folder):
return []
image_paths = []
for ext in ['*.jpg', '*.jpeg', '*.png', '*.JPG', '*.JPEG', '*.PNG']:
image_paths.extend(glob.glob(os.path.join(examples_folder, ext)))
return sorted(image_paths)
def load_example_image(image_path):
"""Load an image from the examples folder"""
if image_path:
return Image.open(image_path)
return None
with gr.Blocks(title="Brain MRI Tumor Classifier", theme=gr.themes.Default()) as demo:
gr.Markdown("""
# 🧠 Brain MRI Tumor Classifier with Occlusion Sensitivity
This application classifies brain MRI images into four categories and visualizes which regions
are most important for the classification decision.
**Classes:** Glioma β€’ Meningioma β€’ No Tumor β€’ Pituitary
""")
with gr.Row():
# Left column - Input controls
with gr.Column(scale=1):
gr.Markdown("### πŸ“ Select or Upload Image")
# Example image selector
example_images = get_example_images('examples/')
if example_images:
example_dropdown = gr.Dropdown(
choices=example_images,
label="Select from Examples folder",
value=None,
interactive=True
)
else:
example_dropdown = gr.Dropdown(
choices=[],
label="No examples found in 'Examples' folder",
interactive=False
)
# OR upload custom image
gr.Markdown("**OR**")
input_image = gr.Image(type="pil", label="Upload Your Own MRI Image")
gr.Markdown("### βš™οΈ Sensitivity Analysis Settings")
patch_size_slider = gr.Slider(
minimum=10, maximum=50, value=50, step=5,
label="Patch Size",
info="Larger = faster but less detailed"
)
stride_slider = gr.Slider(
minimum=5, maximum=20, value=20, step=5,
label="Stride",
info="Larger = faster but less detailed"
)
predict_btn = gr.Button("πŸ” Analyze Image", variant="primary", size="lg")
gr.Markdown("""
---
**ℹ️ Tip:** Select an image from the dropdown or upload your own,
then click 'Analyze Image' to see predictions and sensitivity maps.
---
**πŸ“ Note:** The default values for patch size and stride have been set to the maximum due to computational constraints.
Though the results will be enough for a quick demo.
""")
# Right column - Results
with gr.Column(scale=2):
# Prediction results at the top
output_text = gr.Markdown("### Waiting for analysis...")
gr.Markdown("---")
# Visualization results
gr.Markdown("### πŸ”¬ Sensitivity Analysis Visualizations")
with gr.Row():
output_original = gr.Image(label="Original Image", type="pil")
output_sensitivity = gr.Image(label="Sensitivity Map", type="pil")
with gr.Row():
output_overlay = gr.Image(label="Overlay Visualization", type="pil")
# Event handlers
if example_images:
example_dropdown.change(
fn=load_example_image,
inputs=[example_dropdown],
outputs=[input_image]
)
predict_btn.click(
fn=predict_and_visualize,
inputs=[input_image, patch_size_slider, stride_slider],
outputs=[output_original, output_sensitivity, output_overlay, output_text],
api_name="predict",
)
if __name__ == "__main__":
demo.launch(share=False)