ASD Classifier based on Structural Brain Imaging Data

This project trains a classifier to distinguish Autism Spectrum Disorder (ASD) cases from typical controls using structural MRI features from the ABIDE (Autism Brain Imaging Data Exchange) dataset.

A report containing the aims of this project and results associated with it is available here : report.pdf.

Project structure

.
├── abide -> /mnt/abide/abide/data/Projects/ABIDE_Initiative   # symlink to mounted S3 data
├── misc/                                                       
├── mount_data.sh                                               # mounts the ABIDE S3 bucket
├── report.pdf                                                 
├── requirements.txt
└── src/
    ├── main.py            #  entry point
    ├── data_import.py     # imports raw subject data
    ├── data_handler.py     # loads, harmonizes, and prepares subject data
    ├── classifier.py       # model definition
    ├── train.py            # training loop
    └── analysis.py         # evaluation / results analysis

Setup

Create and activate a virtual environment, then install dependencies:

python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -r requirements.txt

Getting the data

This repository ships with the data already included under misc/, so you can go straight to Usage without fetching anything yourself.

If you'd rather fetch it yourself from the source (for example, to verify provenance or get an updated copy), the data is hosted on a public Amazon S3 bucket and can be mounted locally with s3fs:

1. Install s3fs (e.g. sudo apt install s3fs on Debian/Ubuntu).

2. Create the mount point and mount the bucket:

   sudo mkdir -p /mnt/abide
   ./mount_data.sh
   

   This mounts the public fcp-indi bucket to /mnt/abide and creates a local symlink, abide, pointing into the ABIDE Initiative project folder.

3. Fetch and prepare the data through the CLI:

   cd src
   python main.py fetch
   

   Fetching the full dataset from the bucket can take several hours, which is why a local copy is included in this repo by default.

Usage

All commands are run from the src/ directory:

cd src
python main.py [fetch|process|train|analyse]

Command

Description

fetch

Imports phenotypic and imaging data (from misc/ or the mounted bucket).

process

Loads all subjects, harmonizes the data across sites, and prepares it for training.

train

Trains the classifier.

analyse

Runs evaluation and analysis on the trained model and produces results.

Typical pipeline:

python main.py fetch
python main.py process
python main.py train
python main.py analyse

Requirements

See requirements.txt for the full list of Python dependencies (PyTorch, pandas, etc.).