RetinaScan

Diabetic retinopathy is a leading cause of preventable blindness, and early detection through fundus imaging is the difference-maker. For this project I was awarded a $6,000 UREAP scholarship at TRU to train CNN models that flag retinopathy automatically.

The work covered the full research cycle: literature review, domain-specific preprocessing, model training, benchmarking, and documentation written for reproducibility.

Fundus images need care before a model ever sees them: uneven illumination, lens artifacts, and class imbalance all skew training. The preprocessing pipeline applied domain-specific normalization and augmentation tuned to retinal imagery.

On the modeling side, I applied transfer learning and hybrid architectures, benchmarking seven configurations against consistent evaluation criteria rather than chasing a single headline number.

Every experiment was logged with its data split, preprocessing config, and metrics so results could be reproduced end to end. The benchmarking compared architectures on sensitivity and specificity, the numbers that matter clinically, not just raw accuracy.

A core deliverable was technical documentation ensuring data transparency and reproducibility: dataset provenance, preprocessing decisions, and evaluation protocols are all written down so the next researcher can pick the work up without guessing.

The project delivered a benchmarked set of CNN approaches for retinopathy detection and a documented, reproducible pipeline. It also sharpened the analytical habits I now bring to every ML project: preprocess with domain knowledge, benchmark honestly, document everything.