Critical ML Lab, University of Waterloo

Overview

Self-supervised and active learning for label-efficient medical image segmentation.

Contributions

• Co-authored DiCoH, a self-supervised contrastive pretraining framework for homogeneous medical images (under review), +2.0 mIoU over the strongest baseline at 5% labels
• Developing a reinforcement learning-driven active learning framework for semi-supervised segmentation, leading reward-signal design
• Built a PyTorch active learning pipeline on 80K+ BraTS MRI slices, resolving a class-imbalance failure mode in sample selection

Tools and methods: Python, PyTorch, Self-Supervised Learning, Active Learning, Reinforcement Learning, Semantic Segmentation.