MOST: Multi-formation Soft Masking for Semi-supervised Medical Image Segmentation

Xinyu Liu, Zhen Chen, Yixuan Yuan

Abstract: In semi-supervised medical image segmentation (SSMIS), existing methods typically impose consistency or contrastive regularizations under basic data and network perturbations, and individually segment each voxel/pixel in the image. In fact, a dominating issue in medical scans is the intrinsic ambiguous regions due to unclear boundary and expert variability, whose segmentation requires the information in spatially nearby regions. Thus, these existing works are limited in data variety and tend to overlook the ability of inferring ambiguous regions with contextual information. To this end, we present Multi-Formation Soft Masking (MOST), a simple framework that effectively boosts SSMIS by learning spatial context relations with data regularity conditions. It first applies multi-formation function to enhance the data variety and perturbation space via partitioning and upsampling. Afterwards, each unlabeled data is soft-masked and is constrained to give invariant predictions as the original data. Therefore, the model is encouraged to infer ambiguous regions via varied granularities of contextual information conditions. Despite its simplicity, MOST achieves state-of-the-art performance on four common SSMIS benchmarks. Code and models are released at https://github.com/CUHK-AIM-Group/MOST-SSL4MIS.