AI medical-imaging research pipeline

Alzheimer's Detection Research Pipeline

A reproducible PyTorch pipeline taking raw brain MRI files through preprocessing, training, and cloud staging to evaluation artifacts collaborators can audit.

Overview

Built through the Drew Summer Science Institute at Drew University, this repository moves from raw ADNI and Kaggle MRI files to reproducible preprocessing, EfficientNet-B0 training, Google Cloud staging, evaluation artifacts, and export-ready inference work.

The research question was practical as much as technical: can a compact imaging model distinguish healthy controls from patients on the Alzheimer’s continuum, while keeping the pipeline reproducible enough for collaborators to audit?

This is academic research, not a client product, and no clinical claim is made for it. It appears here because the engineering underneath is the transferable part — a documented data path, configuration-driven runs, GPU training staged on managed cloud infrastructure, and models exported behind an API. That is the same shape as any production machine-learning feature.

Metrics from the research are deliberately not reproduced on this site. They belong with the research write-up and its methodology, not in marketing copy.

Status: University research programme — Drew Summer Science Institute, Summer 2025
Built for: Research collaborators reproducing and auditing imaging experiments
Role: Research engineering, MRI data preparation, model training, evaluation reporting
Timeline: Summer 2025
Platforms: Google Cloud, Local GPU
Categories: AI & Research

Visualisation of the Alzheimer's research MRI processing pipeline

What was built

Responsibilities and core features

Responsibilities

MRI conversion workflow for DICOM and NIfTI inputs
Preprocessing, quality checks, slice extraction, and stratified splits
PyTorch EfficientNet-B0 training with checkpointing and metric artifacts
Google Cloud Storage and Vertex AI staging for GPU training runs
Evaluation reporting and export paths for serving

Core features

Two-notebook workflow: an end-to-end prototype and a cleaner training and evaluation runner
DICOM and NIfTI conversion with PNG slice extraction and quality checks
Stratified 80/10/10 splits
EfficientNet-B0 training with AdamW and cosine learning-rate decay
Timestamped outputs, cloud checkpoints, and repeatable metrics
TorchScript and ONNX export examples with FastAPI serving

Screens

Raw imaging files move through conversion, quality checks, slice extraction, stratified splits, and cloud staging.

Engineering

Problems that came up, and what fixed them

The specifics matter more than the feature list. This is what the work actually consisted of.

Reproducibility as the actual deliverable
The research question was as much practical as technical: keeping the pipeline reproducible enough for collaborators to audit. Python modules and YAML configs carry experiment setup, preprocessing, and run control, so a run can be repeated rather than described.
Documenting the data path as carefully as the model
Medical-imaging work becomes more credible when preprocessing, splits, and outputs are documented as carefully as model code. Conversion, quality checks, slice extraction, and stratified splits are each explicit steps with recorded outputs rather than ad-hoc preparation.

Technical architecture

Jupyter notebooks for an end-to-end prototype and a cleaner day-two training and evaluation runner
Python modules and YAML configs for reusable experiment setup, preprocessing, and run control
MRI conversion workflow for DICOM and NIfTI inputs, PNG slice extraction, quality checks, and stratified splits
PyTorch EfficientNet-B0 training with AdamW, cosine LR decay, checkpointing, and detailed metric artifacts
Google Cloud Storage and Vertex AI staging, with export examples for TorchScript, ONNX, and FastAPI serving

Technology stack

Python
PyTorch
Jupyter
scikit-learn
Google Cloud
Vertex AI
FastAPI

What it adds up to

An end-to-end path from raw MRI files to a trained classifier and evaluation artifacts
Timestamped, repeatable runs with cloud checkpoints
Export-ready inference paths through TorchScript, ONNX, and FastAPI

Related services