Medical image segmentation under multiple real-world constraints
Medical imaging, such as CT scans and MRI, plays a vital role in diagnosing diseases and planning treatments. Computers can help doctors by automatically identifying and outlining regions of interest, such as tumours---a process called image segmentation. However, training these systems for clinical use faces practical challenges: expert annotations are expensive and time-consuming to obtain, and the most informative imaging modalities are often costly or unavailable.
In his thesis, Changtai Li proposes three methods to address these challenges. First, he developed a technique that learns to transform rough, non-expert annotations into expert-quality results. Second, Li introduces a self-supervised approach that enables computers to learn useful patterns from unlabelled 3D scans, reducing the need for extensive expert labelling. Third, Li designed a knowledge transfer framework that allows accurate segmentation using only low-cost images at diagnosis time, even when trained with richer imaging data.
Together, these contributions make medical image segmentation more practical for everyday clinical use by reducing dependence on expensive annotations and costly imaging procedures, while maintaining the accuracy that doctors require.