NeuroFusion AI

• 

  NeuroFusion AI

Inspiration

Alzheimer’s is a devastating disease characterized by a slow onset that is often missed until significant neurological damage has occurred. We were inspired by the Hack4Health mission to democratize computational medicine. We realized that while many models look at MRI scans, few integrate the patient's genetic "blueprint." Our goal was to build a bio-integrated system that cross-references physical brain atrophy with genomic risk markers to catch Alzheimer's earlier than ever before.

What it does

NeuroFusion AI is a multi-modal diagnostic tool. It takes two distinct inputs:

1. Structural MRI Scans: Analyzed via a Deep Convolutional Neural Network to detect hippocampal shrinkage.
2. Genomic Variants (SNPs): Processed through a Multi-Layer Perceptron to calculate polygenic risk scores. The system "fuses" these data streams to produce a Certified Medical AI Report, providing a diagnosis, a confidence score, and actionable clinical precautions.

How we built it

The project was built using PyTorch and Python. We implemented a sophisticated "Late Fusion" architecture:

1. Vision Path: We utilized Transfer Learning with a ResNet-18 backbone. Pre-trained on millions of images, this model is highly sensitive to the subtle structural textures of brain tissue.
2. Genomic Path: We built a custom MLP to process 130 binary-encoded genomic features derived from the ADVP (hg38) dataset.
3. The Pipeline: We engineered a self-healing data loader to handle MRI data stored in Parquet format, converting raw bytes into normalized tensors.

Challenges we ran into

Data Integrity: Our genomic dataset initially contained infinite values and outliers in the P-value and Sample Size columns. We solved this by implementing a robust cleaning pipeline using Median Imputation and Value Clipping.

1. Data Formatting: The MRI dataset used a complex dictionary-based Parquet structure. We had to write a custom wrapper to extract image bytes dynamically and convert them into a format compatible with ResNet-18.

2. Confidence Levels: Initially, the model had low confidence. By implementing Standardization and Batch Normalization, we stabilized the gradients and significantly increased the AI's diagnostic certainty.

Accomplishments that we're proud of

We successfully built a model that doesn't just "predict" but "explains." Our diagnostic output window provides real-world medical advice and precautions based on the AI's findings. We are proud to have moved beyond simple image classification into the realm of Multi-Modal Bio-Integration.

What we learned

We learned that medical data is messy. Scaling P-values and handling NaNs is just as important as the neural network architecture itself. We also gained deep insights into how Residual Connections in ResNet help prevent vanishing gradients in deep medical imaging tasks.

What's next for NeuroFusion AI

We plan to integrate the ADNI (Alzheimer's Disease Neuroimaging Initiative) dataset to further increase our accuracy. We also aim to add a Local Interpretabilty (SHAP) layer to the genomic path, so doctors can see exactly which SNP variants are driving the high-risk score.

Built With

• .npz
• .tsv
• devpost
• devpost-data-formats:-parquet
• google-colab
• matplotlib
• numpy
• pandas
• parquet
• pil
• python
• pytorch
• scikit-learn
• seaborn
• seaborn-tools:-google-colab
• torchvision
• torchvision-libraries:-numpy