Medical Image Processing for Automated Screening Systems

Abstract

This thesis builds a two-stage, fully automated AMD screening pipeline using no expert labels. Stage one generates pseudo-labels by extracting ResNet-18 features, reducing them with PCA, and clustering them into nine AMD subtypes, yielding a balanced set of distribution from a set of 2,914 patient data. Stage two trains a binary ResNet-18 classifier on these pseudo-labels, reaching certain results for Exudative AMD on 419-patient test set.The work is explicitly framed as a proof-of-concept, since all metrics measure agreement with pseudo-labels rather than clinical truths. The main limitation is the lack of expert validation, and the thesis identifies cluster purity review by a retinal specialist as essential for future clinical relevance. Overall, it demonstrates that a weakly supervised AMD classifier can be built from entirely unlabelled OCT data using standard deep learning components.