src/app.py

# app.py
import streamlit as st
import tensorflow as tf
from PIL import Image
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
import cv2
##

# Load saved model (custom CNN)
@st.cache_resource
def load_model():
    """Loads and returns the trained Keras model."""
    model = tf.keras.models.load_model('best_model.h5')
    return model

model = load_model()

# Class labels with descriptions
tumor_info = {
    'glioma': {
        'description': "Glioma is a type of tumor that occurs in the brain and spinal cell. Gliomas begin in the gluey supportive cells (glial cells) that surround nerve cells.",
        'prevalence': "Most common malignant brain tumor in adults",
        'treatment': "Surgery, radiation therapy, chemotherapy"
    },
    'meningioma': {
        'description': "Meningioma is a tumor that arises from the meninges — the membranes that surround the brain and spinal cord. Most meningiomas are non-cancerous (benign).",
        'prevalence': "Most common primary brain tumor (30% of all brain tumors)",
        'treatment': "Monitoring, surgery, radiation therapy"
    },
    'no_tumor': {
        'description': "No signs of tumor detected in the MRI scan. Normal brain tissue appears healthy.",
        'prevalence': "Normal brain MRI",
        'treatment': "No treatment needed"
    },
    'pituitary': {
        'description': "Pituitary tumors are abnormal growths that develop in the pituitary gland. Most are benign and many don't cause symptoms.",
        'prevalence': "10-15% of all primary brain tumors",
        'treatment': "Medication, surgery, radiation therapy"
    }
}

def generate_gradcam(model, img_array, interpolant=0.5):
    """
    Generates Grad-CAM visualization for a custom CNN model.
    Args:
        model: Compiled Keras model.
        img_array: Preprocessed image array (1, 224, 224, 3).
        interpolant: Opacity for heatmap overlay (0-1).
    Returns:
        tuple: (superimposed_img, heatmap) or (None, error_message).
    """
    try:
        # Find the last convolutional layer automatically
        last_conv_layer = None
        for layer in reversed(model.layers):
            if isinstance(layer, tf.keras.layers.Conv2D):
                last_conv_layer = layer
                break
        if last_conv_layer is None:
            raise ValueError("No Conv2D layer found in the model.")

        # Define a symbolic input tensor for the new `gradient_model`.
        grad_model_input = tf.keras.Input(shape=img_array.shape[1:])

        # Reconstruct the forward pass *symbolically* through the original model's layers
        x = grad_model_input
        last_conv_output_symbolic = None
        
        for layer in model.layers:
            x = layer(x)
            if layer == last_conv_layer:
                last_conv_output_symbolic = x

        final_output_symbolic = x

        if last_conv_output_symbolic is None:
            raise ValueError(f"Could not find the symbolic output for the last convolutional layer ('{last_conv_layer.name}').")

        gradient_model = tf.keras.models.Model(
            inputs=grad_model_input,
            outputs=[last_conv_output_symbolic, final_output_symbolic]
        )

        with tf.GradientTape() as tape:
            inputs_for_tape = tf.cast(img_array, tf.float32)
            conv_outputs, predictions = gradient_model(inputs_for_tape)
            
            # Use argmax to get the predicted class index
            pred_index = tf.argmax(predictions[0])
            
            # Extract the loss for the predicted class
            loss = predictions[:, pred_index]

        grads = tape.gradient(loss, conv_outputs)
        
        # --- Crucial Error Handling for Gradients & Heatmap ---
        if grads is None:
            return None, "Grad-CAM failed: Gradients are None. This might indicate an issue with differentiability or an unusual model state for this input."

        # Global average pooling of gradients
        pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2))
        
        # Weight the conv outputs by pooled gradients
        conv_outputs = conv_outputs[0] # Remove batch dimension
        heatmap = tf.reduce_sum(conv_outputs * pooled_grads, axis=-1)
        
        # Normalize the heatmap
        heatmap = tf.maximum(heatmap, 0) # Apply ReLU to heatmap
        
        # Check if heatmap is all zeros AFTER ReLU. If so, normalization will fail.
        max_heatmap_value = tf.math.reduce_max(heatmap)
        if tf.equal(max_heatmap_value, 0):
            return None, "Grad-CAM failed: Heatmap is entirely zero, cannot normalize. This may happen if the model's activations or gradients are all zero for this input and predicted class."

        heatmap = heatmap / max_heatmap_value # Normalize by max value
        heatmap = heatmap.numpy()
        
        # Resize heatmap to original image size
        heatmap = cv2.resize(heatmap, (img_array.shape[2], img_array.shape[1]))
        
        # Convert to RGB heatmap
        heatmap_colored = cv2.applyColorMap(np.uint8(255 * heatmap), cv2.COLORMAP_JET)
        heatmap_colored = cv2.cvtColor(heatmap_colored, cv2.COLOR_BGR2RGB)
        
        # Prepare original image
        img = np.uint8(img_array[0] * 255)
        
        # Superimpose heatmap on original image
        superimposed_img = cv2.addWeighted(
            img, 1 - interpolant,
            heatmap_colored, interpolant,
            0
        )
        
        return superimposed_img, heatmap
    
    except Exception as e:
        return None, f"Grad-CAM failed: {str(e)}"

# --- Streamlit UI (No changes needed below this point) ---
st.set_page_config(
    page_title="Brain Tumor MRI Classifier",
    page_icon="🧠",
    layout="wide",
    initial_sidebar_state="expanded"
)

# Custom CSS
st.markdown("""
<style>
    .reportview-container {
        background: linear-gradient(135deg, #0f2027, #203a43, #2c5364);
        color: white;
    }
    .sidebar .sidebar-content {
        background: #0c151c !important;
    }
    .stButton>button {
        background-color: #4CAF50;
        color: white;
        border-radius: 8px;
        padding: 10px 24px;
        font-weight: bold;
        transition: all 0.3s;
    }
    .stButton>button:hover {
        background-color: #3d8b40;
        transform: scale(1.05);
    }
    .prediction-highlight {
        font-size: 28px;
        font-weight: bold;
        color: #4CAF50;
        text-shadow: 0 0 8px rgba(76, 175, 80, 0.4);
    }
    .stTabs [data-baseweb="tab"] {
        background: rgba(30, 136, 229, 0.2) !important;
        border-radius: 8px !important;
        padding: 10px 20px !important;
        transition: all 0.3s;
    }
    .stTabs [aria-selected="true"] {
        background: #1e88e5 !important;
        font-weight: bold;
    }
</style>
""", unsafe_allow_html=True)

# Title and description
st.title("🧠 Brain Tumor MRI Classification")
st.markdown("This AI-powered tool analyzes brain MRI scans to detect and classify tumors using a Convolutional Neural Network (CNN). Upload an MRI image to get a prediction and detailed insights.")

# Sidebar with info and metrics
with st.sidebar:
    st.header("Model Information")
    st.markdown("""
    - **Model Architecture**: Custom CNN
    - **Training Data**: 1,695 MRI scans
    - **Test Accuracy**: 76.0%
    - **Balanced Accuracy**: 74.8%
    - **Macro F1-Score**: 74.5%
    """)
    st.divider()
    st.header("Performance by Tumor Type")
    with st.expander("Glioma"):
        st.markdown("**Precision**: 0.78 | **Recall**: 0.93 | **F1-Score**: 0.85")
    with st.expander("Meningioma"):
        st.markdown("**Precision**: 0.65 | **Recall**: 0.51 | **F1-Score**: 0.57")
    with st.expander("No Tumor"):
        st.markdown("**Precision**: 0.89 | **Recall**: 0.63 | **F1-Score**: 0.74")
    with st.expander("Pituitary"):
        st.markdown("**Precision**: 0.75 | **Recall**: 0.93 | **F1-Score**: 0.83")
    st.divider()
    st.warning("⚠️ **Disclaimer**: This tool is for educational purposes only. Always consult a medical professional for diagnosis.")

# Main content area
col1, col2 = st.columns([1, 1])

if 'prediction_made' not in st.session_state:
    st.session_state['prediction_made'] = False
    st.session_state['predicted_class'] = None
    st.session_state['confidence'] = None
    st.session_state['gradcam_img'] = None
    st.session_state['heatmap_error'] = None
    st.session_state['prediction_probs'] = None


with col1:
    st.subheader("Upload MRI Scan")
    uploaded_file = st.file_uploader(
        "Choose a brain MRI image (JPEG)", 
        type=["jpg","jpeg"],
        label_visibility="collapsed"
    )

if uploaded_file is not None:
    # --- Single Analysis Block ---
    image = Image.open(uploaded_file).convert('RGB')
    uploaded_file.close()
    img_display = image.copy()
    image = image.resize((224, 224))
    img_array = np.array(image) / 255.0
    img_array = np.expand_dims(img_array, axis=0)

    # Display uploaded image in the second column
    with col2:
        st.subheader("Uploaded MRI Scan")
        st.image(img_display, caption="Original MRI", use_container_width=True)
    
    with st.spinner('Analyzing MRI scan...'):
        prediction = model.predict(img_array, verbose=0) 
        
        predicted_class = list(tumor_info.keys())[np.argmax(prediction)]
        confidence = np.max(prediction) * 100
        
        gradcam_img, heatmap_status = generate_gradcam(model, img_array, interpolant=0.6)

        st.session_state['prediction_made'] = True
        st.session_state['predicted_class'] = predicted_class
        st.session_state['confidence'] = confidence
        st.session_state['gradcam_img'] = gradcam_img
        st.session_state['heatmap_error'] = heatmap_status
        st.session_state['prediction_probs'] = prediction[0]

# --- Results Section (display only if prediction was made) ---
if st.session_state['prediction_made']:
    st.divider()
    
    col_res1, col_res2 = st.columns([1, 2])
      
    with col_res1:
        st.subheader("AI Analysis Result")
        st.markdown(f"<div class='prediction-highlight'>{st.session_state['predicted_class'].replace('_', ' ').title()}</div>", unsafe_allow_html=True)
        
        st.metric("Confidence Level", f"{st.session_state['confidence']:.2f}%")
        st.progress(int(st.session_state['confidence']))
        
        info = tumor_info[st.session_state['predicted_class']]
        with st.expander("Tumor Information", expanded=True):
            st.markdown(f"**Description**: {info['description']}")
            st.markdown(f"**Prevalence**: {info['prevalence']}")
            st.markdown(f"**Treatment**: {info['treatment']}")
        
        st.info("💡 **Clinical Note**: The AI analysis should be reviewed by a qualified radiologist. It is not a substitute for professional medical diagnosis.")

    with col_res2:
        st.subheader("Model Insights")
        
        tab1, tab2, tab3 = st.tabs(["📊 Probability Distribution", "🔥 Attention Map", "📈 Performance Metrics"])
        
        with tab1:
            classes = list(tumor_info.keys())
            probs = st.session_state['prediction_probs'] * 100
            
            fig, ax = plt.subplots(figsize=(10, 5))
            colors = ['#1e88e5' if c != st.session_state['predicted_class'] else '#4CAF50' for c in classes]
            bars = ax.barh([c.replace('_', ' ').title() for c in classes], probs, color=colors)
            ax.set_xlabel('Probability (%)')
            ax.set_title('Prediction Confidence Distribution', fontsize=14, fontweight='bold')
            ax.set_xlim(0, 100)
            
            for bar, prob in zip(bars, probs):
                ax.text(min(prob + 2, 98), bar.get_y() + bar.get_height()/2, f'{prob:.1f}%', 
                                ha='left', va='center', color='white', fontweight='bold', fontsize=12)
            st.pyplot(fig)
        
        with tab2:
            if st.session_state['gradcam_img'] is not None:
                st.image(st.session_state['gradcam_img'], caption="AI Attention Map (Grad-CAM)", use_container_width=True)
                st.markdown("The highlighted areas indicate the regions the model focused on to make its prediction.")
            else:
                st.warning(st.session_state['heatmap_error']) # Display error message from generate_gradcam
        
        with tab3:
            cnn_cm = np.array([
                [74, 6, 0, 0],    # glioma
                [17, 32, 4, 10],  # meningioma
                [1, 10, 31, 7],   # no_tumor
                [3, 1, 0, 50]     # pituitary
            ])
            
            st.write("**Custom CNN Confusion Matrix (Test Set)**")
            fig, ax = plt.subplots(figsize=(8, 6))
            sns.heatmap(
                cnn_cm, annot=True, fmt='d', cmap='Blues', 
                xticklabels=[c.replace('_', ' ').title() for c in tumor_info.keys()],
                yticklabels=[c.replace('_', ' ').title() for c in tumor_info.keys()],
                ax=ax
            )
            ax.set_xlabel('Predicted Label', fontsize=12)
            ax.set_ylabel('True Label', fontsize=12)
            st.pyplot(fig)
            
            st.write("**Performance by Class:**")
            class_data = {
                'Tumor Type': ['Glioma', 'Meningioma', 'No Tumor', 'Pituitary'],
                'Precision': [0.78, 0.65, 0.89, 0.75],
                'Recall': [0.93, 0.51, 0.63, 0.93],
                'F1-Score': [0.85, 0.57, 0.74, 0.83]
            }
            df = pd.DataFrame(class_data).set_index('Tumor Type')
            st.dataframe(df.style.format("{:.2f}").highlight_max(axis=0, color='rgba(76, 175, 80, 0.3)'))

# Footer
st.markdown("---")
st.caption("© 2025 Brain Tumor Classification System | For Research Use Only")