A geometry and saliency driven network with adaptive label refinement for weakly supervised medical image segmentation

Abstract

Weakly supervised medical image segmentation is promising for its low annotation cost and strong performance, with bounding boxes offering notable advantages over image-level and scribble annotations. However, pseudo-labels generated from bounding boxes often suffer from boundary errors and high uncertainty in transition regions between the target and background, affecting segmentation quality. To overcome these challenges, a geometry and saliency driven weakly supervised segmentation network (GSR-Net) is proposed. The saliency optimization and spatial consistency learning module anchors the centers of segmentation targets, forming the basis for subsequent pseudo-label refinement and enhancing overall consistency. The geometry-guided dynamic feature focusing module uses bounding box geometry to create dynamic boundary weights, refining boundary representations and suppressing background interference. Based on the improved localization and refined boundaries, the dynamic propagation and refinement module iteratively optimizes pseudo labels in uncertain regions, further enhancing segmentation accuracy. Additionally, a random expansion and shrinkage strategy for bounding box annotations is introduced to evaluate the model under varied annotation conditions. Experiments on three representative medical image datasets, namely KiTS23, LiTS, and BraTS2021, demonstrate that GSR-Net significantly outperforms existing weakly supervised methods in segmentation accuracy (Dice) and boundary quality (95HD), exhibiting strong generalization in complex scenarios and under weakly supervised conditions.