Skip to content

Data Science

I build generative diffusion models for medical imaging, designed for small clinical datasets and tested against models trained on real data.

Read the paper

Highlights

  • MSc Big Data & Data Science Technology (with Advanced Practice), Northumbria University: Distinction, dissertation marked 81.
  • A two-stage latent diffusion pipeline (a VAEGAN autoencoder and a conditional latent diffusion model with classifier-free guidance) that generates class-conditioned synthetic brain MRI, trained on 295 ADNI participants.
  • A classifier trained only on synthetic scans reached TSTR AUC 0.779 ± 0.031 on 44 held-out real participants, against 0.869 ± 0.027 for the same model trained on real data, a difference not distinguishable at this sample size.
  • No instance memorisation among 880 samples, and a size-matched control exposed a 27.7-percentage-point inflation in the standard nearest-neighbour memorisation metric.
  • Peer-reviewed and published in the Journal of Imaging, 12(9), 426 (MDPI, open access, first author).

Experience

  1. Data and Research Lead

    West London NHS Trust

    Establishing the research programme for a football and mental-health wellbeing service.

  2. Graduate Student Researcher

    Northumbria University, London

    MSc Big Data & Data Science Technology with Distinction. My dissertation, marked 81, developed a conditional latent diffusion model for synthetic brain MRI in Alzheimer's disease.

Selected work

Conditional Brain-MRI Synthesis

MSc dissertation · Generative AI

A two-stage VAEGAN and conditional latent diffusion pipeline that synthesises brain MRI by Alzheimer's disease stage from a 295-subject ADNI cohort, designed for small datasets.

PyTorchMONAIANTsHD-BETDiffusion
Read the paper (DOI)

Publication

Conditional Latent Diffusion for Synthetic Brain MRI in Alzheimer's Disease: A Preprocessing-Focused Pipeline

Fallah, S.; Herzog, N. J. Journal of Imaging, 12(9), 426.

A VAE compressing 256 × 256 coronal slices to a 32 × 32 × 8 latent, then a class-conditional latent diffusion model under classifier-free guidance, trained on 295 ADNI participants. A preprocessing-only ablation improved KID by 0.0147 and precision by 0.069, both 95% intervals excluding zero. Supervised by Dr Nitsa Herzog, Northumbria University.

Peer-reviewed · Open access · First author

Read the paper

Let’s connect