PRJ_006
006
Research · Self-Supervised Learning · Medical AI
Alzheimer's Disease Classification (MSc)
MSc Data Science · University of Greenwich · Dec, 2024 – Dec, 2025
Overview

MSc thesis investigating self-supervised feature learning as a strategy for Alzheimer's Disease classification under data scarcity conditions. The work explores uncertainty-aware prediction frameworks — architectures that don't just classify, but quantify how confident they are in their own outputs.

The Problem

Medical imaging datasets are chronically small and expensive to label. Standard supervised learning struggles in these conditions — models either overfit or fail to generalise across imaging protocols and patient demographics. Self-supervised learning offers a path to useful representations without requiring labelled data at scale. The question this thesis asks is whether those representations are good enough for clinical-grade classification, and how uncertainty should be communicated to clinicians.

Key Metrics
SSL
Learning paradigm
UQ
Uncertainty quantification
MRI
Input modality
2026
Expected completion
Process & Timeline
Stage 1
Literature review
Surveyed self-supervised methods (SimCLR, MoCo, BYOL, MAE) and their applications in medical imaging. Identified gaps in uncertainty-aware approaches for neurodegenerative disease classification.
Stage 2
Pretraining on unlabelled data
Pretrained feature encoder on large unlabelled MRI corpus using contrastive learning. Evaluated representation quality via linear probing against labelled downstream tasks.
Stage 3
Uncertainty framework
Implemented Monte Carlo Dropout and Deep Ensembles for uncertainty estimation. Designed output format that communicates confidence intervals appropriate for clinical context.
Stage 4
Evaluation & writing
Benchmarking against supervised baselines. Ablation studies on pretraining dataset size and augmentation strategies. Thesis write-up in progress.
Tech Stack
PyTorchSimCLRMonte Carlo DropoutDeep EnsemblesMedical Imaging (MRI)Pythonscikit-learnMatplotlib
Critical Self-Evaluation
This is the most intellectually honest work I've done. Self-supervised learning on medical data forces you to confront assumptions baked into standard supervised pipelines — particularly around what "good enough" means when the downstream task has clinical implications. The insight I keep returning to: uncertainty quantification is not a supplementary metric in healthcare AI — it's the primary interface between the model and the clinician. A model that says "I'm 94% confident" and a model that says "I'm 94% confident, but 18% of cases like this fall outside my training distribution" are functionally different tools. I want to keep working at this intersection — where rigorous ML meets contexts where the stakes are genuinely high.