Update app.py

app.py

@@ -9,16 +9,7 @@ import io
 import os
 matplotlib.use('Agg')  # Use non-interactive backend
 
-# Load the model using SavedModel format
-MODEL_PATH = "chest_ct_binary_classifier_densenet_tf_20250427_182239"
-model = tf.saved_model.load(MODEL_PATH)
-infer = model.signatures["serving_default"]  # Get the inference function
-
-# Get input and output tensor names
-input_tensor_name = list(infer.structured_input_signature[1].keys())[0]
-output_tensor_name = list(infer.structured_outputs.keys())[0]
-
-# Image size - matching what your model was trained on
+# Image size - matching what the model was trained on
 IMG_SIZE = 256
 
 # Function for preprocessing
@@ -26,29 +17,15 @@ def preprocess_image(image):
     img = Image.fromarray(image).convert('RGB')
     img = img.resize((IMG_SIZE, IMG_SIZE))
     img_array = np.array(img) / 255.0
-    return np.expand_dims(img_array, axis=0).astype(np.float32)  # Cast to float32 for TF
+    return np.expand_dims(img_array, axis=0)
 
-# Make prediction with the SavedModel
-def predict_with_saved_model(image_tensor):
-    # Create the input tensor with the right name
-    input_dict = {input_tensor_name: image_tensor}
-    # Run inference
-    output = infer(**input_dict)
-    # Get the prediction value
-    prediction = output[output_tensor_name].numpy()[0][0]
-    return prediction
-
-# Generate Grad-CAM using the SavedModel
-# Note: Grad-CAM is more complex with SavedModel format, so we'll use a simplified approach
+# Generate attention map using edge detection (simplified)
 def generate_attention_map(img_array, prediction):
-    # Since getting Grad-CAM from SavedModel is complex, let's use a simplified heatmap
-    # This is a placeholder - in production you may want to implement a proper CAM
-    
-    # For demo purposes, we'll create a simple attention map based on image features
+    # Convert to grayscale
     gray = cv2.cvtColor(img_array[0].astype(np.float32), cv2.COLOR_RGB2GRAY)
     blur = cv2.GaussianBlur(gray, (5, 5), 0)
     
-    # Use simple edge detection as a proxy for "interesting" regions
+    # Use edge detection to find interesting regions
     sobelx = cv2.Sobel(blur, cv2.CV_64F, 1, 0, ksize=3)
     sobely = cv2.Sobel(blur, cv2.CV_64F, 0, 1, ksize=3)
     magnitude = np.sqrt(sobelx**2 + sobely**2)
@@ -56,7 +33,7 @@ def generate_attention_map(img_array, prediction):
     # Normalize to 0-1
     magnitude = (magnitude - magnitude.min()) / (magnitude.max() - magnitude.min() + 1e-8)
     
-    # Apply sigmoid weighting based on prediction (higher probability = more intensity)
+    # Weight by prediction confidence
     weight = 0.5 + (prediction - 0.5) * 0.5  # Scale between 0.5-1 based on prediction
     magnitude = magnitude * weight
     
@@ -74,8 +51,9 @@ def predict_and_explain(image):
     # Preprocess the image
     preprocessed = preprocess_image(image)
     
-    # Make prediction
-    prediction = predict_with_saved_model(preprocessed)
+    # For demo, use a fixed prediction
+    # In a real deployment, you would load and use your model
+    prediction = 0.75  # Simulated cancer probability
     
     # Generate attention map
     heatmap, attention = generate_attention_map(preprocessed, prediction)
@@ -142,7 +120,7 @@ with gr.Blocks(title="Chest CT Scan Cancer Detection") as demo:
     - Middle: Feature map highlighting areas with distinctive patterns
     - Right: Overlay of the feature map on the original image
     
-    The model was trained on a dataset of chest CT scans containing normal images and various types of lung cancer (adenocarcinoma, squamous cell carcinoma, and large cell carcinoma).
+    Note: This is a demonstration version without the full model due to size limitations.
     """)
     
     submit_btn.click(