AI for Medical Imaging & Research

The problem

Early detection of neurodegenerative disease and accurate tumour classification run into the same wall: a single imaging modality does not carry enough signal. A radiologist reading an MRI is combining it with everything else they know about the patient. A model that sees only the scan is working with less information than the clinician it is meant to support.

What I built

Two lines of work. For Alzheimer detection, models that read the MRI as a volume rather than as a stack of slices — slicing away the spatial relationships throws out exactly the signal you came for. For tumour classification, a pipeline that fuses imaging with genomic profiles, with each branch trainable on its own so it is possible to say whether fusion actually helped or one source was carrying the result alone.

Explainability was a requirement from the first line rather than a later addition. A clinician will not act on a number without a reason, and retrofitting the reason usually means changing the model anyway, later, under worse conditions.

Where it stands

The work led to a co-authored patent on the diagnostic methodology, alongside contributed research. The habit it left me with travels to every other domain: a result you cannot explain is not a result you can deploy.

What I did

Developed deep learning models for Alzheimer detection from 3D brain MRI
Built a multi-modal pipeline fusing imaging features with genomic markers for tumour classification
Handled volumetric preprocessing: normalisation, registration and the class imbalance that comes with clinical data
Designed the branches to be evaluable independently, so a gain could be attributed to a modality
Co-authored a patent with the research group on the diagnostic methodology

Impact

Patent co-authored and contributed research
Explainability treated as a requirement, not a paper section

Stack

Python
PyTorch

3D CNN
Multi-modal
Medical AI