app.py

"""

Chest X-ray Classification App with Multiple Models

Integrates multiple models with unified Grad-CAM visualization

"""

import os
import torch
import numpy as np
import gradio as gr
import torchxrayvision as xrv
from PIL import Image
import torchvision.transforms as transforms
from torchvision import models
import torch.nn as nn
import numpy
import torch.serialization
import matplotlib.pyplot as plt
import cv2
import matplotlib
matplotlib.use('Agg')  # Use non-interactive backend

# Import the custom EfficientNet-B3 model from separate module
# This is actually the DannyNet model but renamed for consistency
import efficientnet_b3_custom

torch.serialization.add_safe_globals([numpy.core.multiarray.scalar])

# Define the available models - removed parenthetical descriptors
MODELS = {
    "DenseNet121": "densenet121-res224-nih",
    "EfficientNet-B3": "efficientnet_b3_custom",  # This is the DannyNet model
    "EfficientNet-B3 O": "efficientnet_b3",
    "EfficientNet-B0": "efficientnet_b0"
}

# NIH ChestX-ray14 pathologies
NIH_PATHOLOGIES = [
    'Atelectasis', 'Cardiomegaly', 'Effusion', 'Infiltration',
    'Mass', 'Nodule', 'Pneumonia', 'Pneumothorax', 'Consolidation',
    'Edema', 'Emphysema', 'Fibrosis', 'Pleural_Thickening', 'Hernia'
]

# Cache for loaded models
loaded_models = {}

# For Grad-CAM
activation = {}
gradient = {}


# Hook functions for Grad-CAM
def get_activation(name):
    def hook(model, input, output):
        activation[name] = output  # Keep gradient tracking
    return hook


def get_gradient(name):
    def hook(grad):
        gradient[name] = grad
    return hook


def load_model(model_name):
    if model_name in loaded_models:
        return loaded_models[model_name]

    model_type = MODELS[model_name]

    try:
        if model_name == "DenseNet121":
            # Load DenseNet121 from TorchXRayVision
            model = xrv.models.DenseNet(weights="densenet121-res224-nih")
            model.eval()
            loaded_models[model_name] = model
            return model

        elif model_type == "efficientnet_b3_custom":
            # Load the custom EfficientNet-B3 model (actually DannyNet)
            model_path = "dannynet-55-best_model_20250422-211522.pth"
            if os.path.exists(model_path):
                model = efficientnet_b3_custom.load_model(model_path, device='cpu')
                print(f"Successfully loaded EfficientNet-B3 from {model_path}")
                loaded_models[model_name] = model
                return model
            else:
                print(f"Model file not found: {model_path}")
                print("Please place the model file in the same directory as this script.")
                return None

        elif model_type == "efficientnet_b3":
            # Import EfficientNet dynamically to avoid dependency issues
            try:
                from efficientnet_pytorch import EfficientNet
                model = EfficientNet.from_name('efficientnet-b3', num_classes=14)
            except ImportError:
                # Fallback to torchvision if efficientnet_pytorch is not available
                model = models.efficientnet_b3(pretrained=False)
                num_ftrs = model.classifier[1].in_features
                model.classifier[1] = nn.Linear(num_ftrs, 14)

            # Load your trained weights
            model_path = os.path.join("weights", "best_model_b3.pt")
            if os.path.exists(model_path):
                # Explicitly set weights_only=False for PyTorch 2.6+ compatibility
                checkpoint = torch.load(model_path, map_location='cpu', weights_only=False)
                if "model_state_dict" in checkpoint:
                    model.load_state_dict(checkpoint["model_state_dict"], strict=False)
                else:
                    model.load_state_dict(checkpoint, strict=False)

                print(f"Successfully loaded EfficientNet-B3 Original from {model_path}")
            else:
                print(f"Model file not found: {model_path}")
                print("Using a new model instance. Please place your trained model in the weights directory.")

        elif model_type == "efficientnet_b0":
            # Import EfficientNet dynamically to avoid dependency issues
            try:
                from efficientnet_pytorch import EfficientNet
                model = EfficientNet.from_name('efficientnet-b0', num_classes=14)
            except ImportError:
                # Fallback to torchvision if efficientnet_pytorch is not available
                model = models.efficientnet_b0(pretrained=False)
                num_ftrs = model.classifier[1].in_features
                model.classifier[1] = nn.Linear(num_ftrs, 14)

            # Load your trained weights
            model_path = os.path.join("weights", "best_model_b0.pt")
            if os.path.exists(model_path):
                # Explicitly set weights_only=False for PyTorch 2.6+ compatibility
                checkpoint = torch.load(model_path, map_location='cpu', weights_only=False)
                if "model_state_dict" in checkpoint:
                    model.load_state_dict(checkpoint["model_state_dict"], strict=False)
                else:
                    model.load_state_dict(checkpoint, strict=False)

                print(f"Successfully loaded EfficientNet-B0 from {model_path}")
            else:
                print(f"Model file not found: {model_path}")
                print("Using a new model instance. Please place your trained model in the weights directory.")

        model.eval()
        loaded_models[model_name] = model
        return model

    except Exception as e:
        print(f"Error loading model {model_name}: {e}")
        return None


def preprocess_image_densenet(img):
    """Preprocess an image for the DenseNet model."""
    # Convert to grayscale if it's a color image
    if len(img.shape) > 2:
        img = img.mean(2)

    # Normalize the image
    img = xrv.datasets.normalize(img, 255)

    # Add channel dimension
    if len(img.shape) == 2:
        img = img[None, ...]

    return img


def preprocess_image_efficientnet(img, img_size=224):
    """Preprocess an image for the EfficientNet models."""
    # Convert to PIL Image if it's a numpy array
    if isinstance(img, np.ndarray):
        img = Image.fromarray(img)

    # Ensure image is in RGB mode
    img = img.convert('RGB')

    # Define preprocessing transforms
    transform = transforms.Compose([
        transforms.Resize((img_size, img_size)),
        transforms.CenterCrop(img_size),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])

    # Apply transforms
    img_tensor = transform(img)

    return img_tensor


# Unified Grad-CAM implementation for all models
def compute_unified_gradcam(model, model_name, img_tensor, target_class_idx):
    """

    Compute Grad-CAM using a unified approach for all models

    This ensures consistent visualization style across all models

    """
    # For the custom EfficientNet-B3 model (DannyNet), use its dedicated function
    if MODELS[model_name] == "efficientnet_b3_custom":
        return efficientnet_b3_custom.compute_gradcam(model, img_tensor, target_class_idx)

    # For other models, implement a similar approach to ensure consistent results
    # Register hooks
    activation.clear()
    gradient.clear()

    # Select appropriate target layer based on model type
    if model_name == "DenseNet121":
        target_layer = model.features.denseblock3
    elif "EfficientNet" in model_name:
        if hasattr(model, '_blocks'):
            middle_idx = len(model._blocks) // 2
            target_layer = model._blocks[middle_idx]
        else:
            middle_idx = len(model.features) // 2
            target_layer = model.features[middle_idx]
    else:
        # Default to last feature layer
        target_layer = model.features[-1]

    # Register forward hook
    handle = target_layer.register_forward_hook(get_activation('target_layer'))

    # Ensure input tensor requires gradients
    img_tensor_for_gradcam = img_tensor.clone().requires_grad_(True)

    # Forward pass
    model.zero_grad()

    # Handle different model output formats
    if model_name == "DenseNet121":
        output = model(img_tensor_for_gradcam.unsqueeze(0))
    else:
        output = model(img_tensor_for_gradcam.unsqueeze(0))
        output = torch.sigmoid(output)

    # Target for backprop
    if target_class_idx is not None:
        score = output[0, target_class_idx]
    else:
        score, _ = output.max(dim=1)
        score = score[0]

    # Backward pass with retain_graph to avoid errors
    score.backward(retain_graph=True)

    # Clean up hook
    handle.remove()

    # Get activations
    if 'target_layer' not in activation:
        print("No activation captured")
        return None

    activations = activation['target_layer']

    # Try different approaches to get gradients
    try:
        # Get gradients using autograd
        gradients = torch.autograd.grad(score, activations,
                                       create_graph=True, retain_graph=True)[0]
    except Exception as e:
        print(f"Gradient calculation failed: {e}")
        # Create dummy gradients as fallback
        gradients = torch.ones_like(activations)

    # Use global average pooling with absolute values for better feature highlighting
    pooled_gradients = torch.mean(torch.abs(gradients), dim=[0, 2, 3])

    # Weight activation maps with gradients
    for i in range(activations.size(1)):
        activations[:, i, :, :] *= pooled_gradients[i]

    # Sum along channels for final heatmap
    heatmap = torch.sum(activations, dim=1).squeeze().cpu().detach().numpy()

    # ReLU on the heatmap
    heatmap = np.maximum(heatmap, 0)

    # Apply gamma correction to enhance contrast
    gamma = 0.7  # Values less than 1 enhance bright regions
    heatmap = np.power(heatmap, gamma)

    # Normalize heatmap
    if np.max(heatmap) > 0:
        heatmap = heatmap / np.max(heatmap)

    # Apply threshold to remove noise
    threshold = 0.2  # Only keep values above 20% of max
    heatmap[heatmap < threshold] = 0

    # Re-normalize after thresholding
    if np.max(heatmap) > 0:
        heatmap = heatmap / np.max(heatmap)

    # Resize to 224x224
    heatmap = cv2.resize(heatmap, (224, 224))

    return heatmap


# Unified Grad-CAM overlay function for all models
def apply_unified_gradcam(original_img, heatmap, alpha=0.6):
    """

    Apply Grad-CAM heatmap to the original image using a unified approach

    This ensures consistent visualization style across all models

    """
    # Convert to numpy if it's a PIL Image
    if isinstance(original_img, Image.Image):
        original_img = np.array(original_img)

    # Resize original image to 224x224
    original_img = cv2.resize(original_img, (224, 224))

    # Convert original image to RGB if it's grayscale
    if len(original_img.shape) == 2:
        original_img = np.stack([original_img] * 3, axis=2)
    elif len(original_img.shape) == 3 and original_img.shape[2] == 1:
        original_img = np.concatenate([original_img] * 3, axis=2)

    # Convert heatmap to uint8 before applying median blur
    heatmap_uint8 = np.uint8(heatmap * 255)
    heatmap_blurred = cv2.medianBlur(heatmap_uint8, 7)
    # Convert back to float in range [0,1]
    heatmap = heatmap_blurred.astype(float) / 255.0

    # Apply colormap to heatmap - Use COLORMAP_HOT for better medical visualization
    heatmap_colored = cv2.applyColorMap(np.uint8(255 * heatmap), cv2.COLORMAP_HOT)

    # Convert to RGB if needed
    if len(original_img.shape) == 3 and original_img.shape[2] == 3:
        heatmap_colored = cv2.cvtColor(heatmap_colored, cv2.COLOR_BGR2RGB)

    # Create a copy of the original image for overlay
    original_img_float = original_img.astype(float)

    # Superimpose heatmap on original image
    superimposed_img = heatmap_colored * alpha + original_img_float * (1 - alpha * 0.5)
    superimposed_img = np.clip(superimposed_img, 0, 255).astype(np.uint8)

    # Add contour lines for the most significant regions
    binary_heatmap = (heatmap > 0.5).astype(np.uint8) * 255
    contours, _ = cv2.findContours(binary_heatmap, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    cv2.drawContours(superimposed_img, contours, -1, (255, 255, 255), 1)

    return superimposed_img


def predict_with_gradcam(image, model_name, confidence_threshold=0.5):
    """Make predictions and generate Grad-CAM visualization."""
    if image is None:
        return None, "Please upload an image."

    try:
        # Create weights directory if it doesn't exist (for EfficientNet models)
        if "EfficientNet" in model_name:
            os.makedirs("weights", exist_ok=True)

        # Load the model
        model = load_model(model_name)

        if model is None:
            return None, f"Failed to load model {model_name}. Please check the console for details."

        # Save original image for visualization
        original_img = np.array(image).copy()

        # Process based on model type
        if model_name == "DenseNet121":
            # Read and preprocess the image for DenseNet from TorchXRayVision
            img = np.array(Image.fromarray(image).convert('RGB'))
            img_processed = preprocess_image_densenet(img)

            # Create transforms
            transform = transforms.Compose([
                xrv.datasets.XRayCenterCrop(),
                xrv.datasets.XRayResizer(224)
            ])

            # Apply transforms
            img_processed = transform(img_processed)

            # Convert to tensor
            img_tensor = torch.from_numpy(img_processed)

            # Make prediction
            with torch.no_grad():
                output = model(img_tensor.unsqueeze(0))
                probabilities = output.squeeze().numpy()

            # Create dictionary of results
            results = {pathology: float(prob) for pathology, prob in zip(NIH_PATHOLOGIES, probabilities)}

        elif MODELS[model_name] == "efficientnet_b3_custom":
            # Use the custom EfficientNet-B3 module for preprocessing and prediction
            img_tensor = efficientnet_b3_custom.preprocess_image(image)

            # Make prediction
            results = efficientnet_b3_custom.predict(model, img_tensor)

        else:  # Other EfficientNet models
            # Preprocess the image for EfficientNet
            img_tensor = preprocess_image_efficientnet(image)

            # Make prediction
            with torch.no_grad():
                output = model(img_tensor.unsqueeze(0))
                probabilities = torch.sigmoid(output).squeeze().numpy()

            # Create dictionary of results
            results = {pathology: float(prob) for pathology, prob in zip(NIH_PATHOLOGIES, probabilities)}

        # Sort results by probability (descending)
        sorted_results = dict(sorted(results.items(), key=lambda item: item[1], reverse=True))

        # Get top pathologies above threshold
        top_pathologies = [p for p, prob in sorted_results.items() if prob >= confidence_threshold]

        # Generate Grad-CAM for top pathologies
        gradcam_img = None

        if top_pathologies:
            # Get index of top pathology
            if model_name == "DenseNet121" or "EfficientNet" in model_name:
                top_pathology = top_pathologies[0]
                target_idx = NIH_PATHOLOGIES.index(top_pathology)

            # Compute Grad-CAM using unified approach
            heatmap = compute_unified_gradcam(model, model_name, img_tensor, target_idx)

            if heatmap is not None:
                # Apply Grad-CAM overlay using unified approach
                gradcam_img = apply_unified_gradcam(original_img, heatmap)

        # If no Grad-CAM was generated, return the original image
        if gradcam_img is None:
            gradcam_img = original_img

        # Format results for display
        result_html = "<h3>Detected Conditions:</h3><ul>"
        detected_count = 0

        for pathology, prob in sorted_results.items():
            if prob >= confidence_threshold:
                detected_count += 1
                result_html += f"<li><b>{pathology}</b>: {prob:.4f} ({prob * 100:.1f}%)</li>"

        if detected_count == 0:
            result_html += "<li>No conditions detected above the confidence threshold.</li>"

        result_html += "</ul>"

        # Add a section for all probabilities
        result_html += "<h3>All Probabilities:</h3><ul>"
        for pathology, prob in sorted_results.items():
            if pathology in top_pathologies:
                result_html += f"<li>{pathology}: {prob:.4f} ({prob * 100:.1f}%) - <span style='color:red'>Used for Grad-CAM</span></li>"
            else:
                result_html += f"<li>{pathology}: {prob:.4f} ({prob * 100:.1f}%)</li>"
        result_html += "</ul>"

        # Add explanation about Grad-CAM
        result_html += "<h3>About Grad-CAM:</h3>"
        result_html += "<p>Grad-CAM highlights regions that influenced the model's prediction for the detected conditions. "
        result_html += "Red/yellow areas indicate regions of high importance for the diagnosis.</p>"

        # Add explanation about the improved visualization
        result_html += "<p><b>Improved Visualization:</b> This enhanced Grad-CAM uses medical imaging-specific techniques to better highlight clinically relevant regions. "
        if len(top_pathologies) > 1:
            result_html += f"The visualization combines information from {len(top_pathologies)} detected conditions for a more comprehensive view.</p>"
        else:
            result_html += "The visualization focuses on the most significant condition detected in the X-ray.</p>"

        return gradcam_img, result_html

    except Exception as e:
        import traceback
        traceback_str = traceback.format_exc()
        print(f"Error processing image: {str(e)}")
        print(traceback_str)
        return None, f"Error processing image: {str(e)}"


# Create the Gradio interface
with gr.Blocks(title="Chest X-ray Disease Classifier with Improved Grad-CAM") as demo:
    gr.Markdown("# Chest X-ray Disease Classifier with Improved Grad-CAM")
    gr.Markdown("Upload a chest X-ray image and select a model to detect potential conditions. The enhanced Grad-CAM visualization will highlight clinically relevant regions influencing the diagnosis.")

    with gr.Row():
        with gr.Column(scale=1):
            input_image = gr.Image(label="Upload Chest X-ray Image", type="numpy")
            model_dropdown = gr.Dropdown(
                choices=list(MODELS.keys()),
                value="EfficientNet-B3",  # Default to the DannyNet model
                label="Select Model"
            )
            confidence = gr.Slider(
                minimum=0.0,
                maximum=1.0,
                value=0.5,
                step=0.05,
                label="Confidence Threshold"
            )
            submit_button = gr.Button("Analyze X-ray")

        with gr.Column(scale=2):
            # REMOVED: Removed the original image display as requested by the user
            gradcam_image = gr.Image(label="Grad-CAM Visualization")
            output_text = gr.HTML(label="Results")

    submit_button.click(
        fn=predict_with_gradcam,
        inputs=[input_image, model_dropdown, confidence],
        outputs=[gradcam_image, output_text]
    )


# Launch the app
if __name__ == "__main__":
    demo.launch()