Generative AI · medical imaging

Data Science

Generative diffusion models for medical imaging, built to hold up on small clinical datasets — and measured without flattering myself.

Data Science — the essentials

  • MSc Big Data & Data Science Technology (with Advanced Practice), Northumbria — Distinction, dissertation marked 81.
  • Two-stage latent-diffusion pipeline (VAEGAN autoencoder + conditional LDM with classifier-free guidance) generating class-conditioned synthetic brain MRI from 295 ADNI subjects.
  • A classifier trained only on synthetic scans reached TSTR AUC 0.754 vs a 0.810 real-data baseline — ~93% of the diagnostic signal retained.
  • Caught and corrected a 33-percentage-point set-size confound in the standard nearest-neighbour memorisation metric.
  • Paper in preparation for MDPI's Journal of Imaging. Research Assistant (volunteer), West London NHS Trust.

Selected work

What I've built and studied

MSc dissertation · Generative AI

Conditional Brain-MRI Synthesis

A two-stage VAEGAN and conditional latent-diffusion pipeline synthesising Alzheimer's-stage brain MRI from a 295-subject ADNI cohort — built to hold up on the small datasets where most methods fall apart.

PyTorchMONAIANTsHD-BETDiffusion
Paper in preparation (J. Imaging)

Research

Publication & dissertation

MSc dissertation · 2026

Conditional Synthesis of Brain MRI for Alzheimer's Disease Research Using Latent Diffusion Models

A two-stage VAEGAN and conditional LDM pipeline on 295 ADNI subjects, built around careful preprocessing. Beyond the generative results, the work caught and corrected a 33-percentage-point set-size confound in a standard memorisation metric. Supervised by Dr Nitsa Herzog, Northumbria University.

Paper in preparation for MDPI's Journal of Imaging.

Contact

Let's talk

Emailcontact@soheilfallah.com
LinkedInlinkedin.com/in/soheilfallah
GitHubgithub.com/soheilfallah
ResearchGateSoheil Fallah