In low-resource clinical settings, diagnostic accuracy for cervical cancer screening depends heavily on clinician experience — creating wide variability in outcomes for a condition that is highly treatable when caught early. This project applied computer vision to Visual Inspection with Acetic Acid (VIA), a low-cost screening technique widely used across LMICs, to reduce that variability with an AI-assisted pipeline. As team lead and ML engineer, I owned technical direction and coordination across preprocessing, segmentation, and classification components, including code review across the full pipeline.
My primary technical focus was the backbone pretraining strategy — specifically implementing SimCLR, a self-supervised contrastive learning approach, to improve feature representations before downstream tasks and address the challenge of limited labeled data. I also supported integration of pretrained backbones into classification and segmentation models and analyzed failure cases to guide architectural adjustments. The project demonstrated meaningful downstream performance gains from thoughtful model initialization under real-world data constraints. GitHub