Scribble-Based Medical Image Segmentation via Shape Perturbation Consistency and Boundary Enhancement Constraint

Acquiring extensive, fully annotated datasets is a labor-intensive and costly process, especially in the medical domain. Scribble annotation offers a more economical alternative, significantly reducing the annotation costs associated with full annotation. However, training segmentation networks from scribbles remains challenging due to incomplete target shape and boundary information in scribble supervision. In this work, we propose a novel scribble-guided weakly supervised medical image segmentation framework, comprising the shape perturbation consistency (SPC) module and the boundary enhancement constraint (BEC) module. Specifically, the SPC module is designed to extract both local and global shape priors through jigsaw puzzle augmentation and intensity augmentation from the images. Furthermore, the BEC module employs the edges extracted from the original image and the pseudo-labels generated by the network’s predictions as supervision, effectively imposing explicit boundary constraints. Experimental results on two public datasets, ACDC and MSCMRseg, demonstrate that our method significantly outperforms state-of-the-art weakly supervised learning approaches. The code is available at https://github.com/SX-SS/SPNet.