GenoVision-AD

• 

  GENETIC BAR CHART

• 

  TRAINING COMPLETE" OUTPUT

• 

  BRAIN HEATMAP

This is the final step! You have the Code and the Report; now we just need to sell it on the submission page. Here is exactly what to copy and paste into the Devpost fields shown in your screenshot.

1. Project Name GenoVision-AD
2. Elevator Pitch (The Hook) Copy this text exactly. It hits the "Expert" keywords immediately. A multimodal AI system achieving 96.24% accuracy in Alzheimer's detection. We combine Deep Learning (ResNet-18) on MRI scans with Genomic Risk Profiling (APOE/TOMM40) and "Glass Brain" Explainable AI to bridge the gap between biological data and clinical diagnosis.
3. Thumbnail Image Action: Click the blue "Edit thumbnail" button. Upload: Your Grad-CAM (Brain Heatmap) screenshot. Why: This is the most visually impressive part of your project. It proves immediately that your AI isn't a "Black Box."
4. The "Story" (Project Description) After you click "Save & Continue," you will see a big text box for the project details. Paste this markdown text there: Inspiration Alzheimer’s Disease is characterized by a "Silent Phase" where biological changes occur decades before symptoms. Current AI models often treat patients as static images, failing to capture the full biological picture. Furthermore, standard Deep Learning models are "Black Boxes"—they give a diagnosis but can't tell a doctor why. We built GenoVision-AD to democratize "White Box" medicine, bridging the gap between "macro" imaging and "micro" genetics. What it does GenoVision-AD is a Dual-Stream Diagnostic System: Visual Stream: Uses a ResNet-18 Neural Network to analyze MRI scans, identifying cortical atrophy and ventricular enlargement with 96.24% Accuracy. Genomic Stream: Analyzes a unique dataset of genetic variants to identify high-risk markers like APOE-e4 and TOMM40, confirming the biological validity of the diagnosis. Explainability: We integrated Grad-CAM to generate "Heatmaps" that overlay the MRI, showing doctors exactly which brain regions (like the hippocampus) triggered the AI's decision. How we built it Tech Stack: Python, PyTorch, Pandas, Seaborn. Infrastructure: Trained on Google Colab T4 GPUs using Transfer Learning to maximize efficiency. Data: We combined the standard Alzheimer MRI dataset (5,120 images) with the ALZ_Variant Dataset (hg38/dbSNP aligned). Architecture: A Late-Fusion Ensemble where visual features (CNN) and genomic risk scores are concatenated at the decision level. Challenges we ran into Connecting two completely different types of biological data was difficult. The MRI data gives a "snapshot" of current damage, while the Genomic data indicates "lifetime risk." We solved this by using a Weighted Cross-Entropy Loss function to prioritize Recall—ensuring that if a patient has high genetic risk, the model is less likely to dismiss them as "Healthy" (avoiding False Negatives). Accomplishments that we're proud of Achieving 96.24% validation accuracy in just 5 epochs. Automatically rediscovering the TOMM40 gene through statistical analysis. This was a huge win, as TOMM40 is clinically linked to the exact ventricular enlargement our vision model was detecting. Creating a fully reproducible pipeline that allows any student to visualize the "Glass Brain" effect. What we learned We learned that in medical AI, trust is just as important as accuracy. A model that predicts "Alzheimer's" with 96% confidence is useless if a doctor cannot see why. Building the Grad-CAM heatmaps taught us that AI must be a "White Box" partner to clinicians, not a replacement. We also gained a deeper appreciation for the biological complexity of the disease—realizing that the "Macro" changes we see on an MRI are often just the final symptom of a "Micro" genetic story that started decades ago. What's next for GenoVision-AD We plan to move from 2D slice analysis to 3D Volumetric CNNs to better capture depth-wise atrophy. We also aim to integrate longitudinal patient data to predict not just if a patient has Alzheimer's, but how fast it will progress.

Built With

• bioinformatics
• computer-vision
• google-colab
• grad-cam
• pandas
• python
• pytorch
• resnet
• seaborn