Leak-Free Alzheimer MRI Classification + Explainable Genetic Risk Insights

A reproducible prediction system for Alzheimer’s dementia stage classification on MRI scans using a leak-free data pipeline, powered by a strong EfficientNet CNN backbone, and enhanced with Grad-CAM spatial attention maps and independent genetic risk analysis, fused only at inference for clinician-aligned explainability.

---

🚀 Features that make this project stand out

• 🧪 Zero Data Leakage → Filename-level dedup + fixed 70/15/15 stratified split
• 🧠 EfficientNet CNN Backbone → Robust MRI feature learning
• 🔥 Grad-CAM Heatmaps (Smooth & Anatomical) → True spatial attention, no pooling artifacts
• 🧬 Genetic Risk Inference (Independent Model) → Odds Ratio + −log10(p-value) scoring
• 🧩 Late Fusion at Inference Only → No invalid MRI-genetics joint training
• 📊 Competition-Grade Metrics → Accuracy, Macro-F1, AUC (OVR), Confidence Analysis
• 💻 Reproducible Notebook → Cloud GPU support (cuda:0)

---

🧠 Grad-CAM Visualization Samples

Panel Output	Description
Original MRI → Grad-CAM → Overlay	Shows exactly where model focuses in the brain

These attention maps are extracted from the last convolutional layer before pooling, resized smoothly to 224×224, and inverted before JET colormap for clinician-friendly “hot attention” visuals.

---

📊 Separate Model Performance

🧠 MRI Model Results

🧬 Genetic Model Results (Independent)

Both models are trained and evaluated separately never jointly to avoid invalid learning or leakage.