Annotation-efficient and reliable medical image segmentation

Abstract

Medical image segmentation, which automatically delineates anatomical structures or lesions, is a key step for diagnosis, surgical planning, and disease monitoring. While deep learning based algorithms have delivered state-of-the-art segmentation performance, their integration into the clinical workflow is hindered by two major challenges: the high cost of acquiring large annotated datasets and the lack of guarantees on model reliability. The objective of this thesis is to develop more reliable and annotation-efficient segmentation algorithms in order to facilitate their clinical adoption. In particular, we address each limitation by targeting a distinct stage of the model development life-cycle. First, we propose a stochastic batch active learning framework that computes uncertainty at the batch level to strategically select, before training, the most informative samples to annotate from large unlabeled datasets. By implicitly enforcing diversity in the selection without additional computational cost, our strategy achieves better segmentation performance than purely uncertainty-based and random sampling, given a fixed data sampling budget. Second, we introduce a weakly-supervised prompt learning framework that adapts large promptable vision foundation models to medical tasks using only bounding box annotations. By applying box-based spatial constraints and consistency regularization to compensate for the reduced label information, our approach avoids costly pixel-level supervision during training and enables resource-efficient segmentation. Results show that weakly-supervised prompt learning is a scalable alternative to fully-supervised specialization of both foundation and non-foundation models. Third, we develop an anatomically-aware conformal prediction framework to provide statistical guarantees on the segmentation outputs after deployment. Specifically, we draw on the dense feature representations of vision foundation models to integrate anatomical context into the conformal process and construct geometrically consistent and statistically valid prediction sets. Our framework can be appended to any trained segmentation model without retraining, making it broadly applicable across architectures and clinical tasks. Evaluated across multiple medical imaging modalities and anatomical targets, these contributions bring medical image segmentation closer to clinical deployment by reducing the annotation burden and providing reliability guarantees.

Date	2 Jun 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Hervé Lombaert (Supervisor) & Christian Desrosiers (Co-supervisor)