MoST-SAM: A Multimodal Self-Training Framework for Annotation-Efficient Basal Cell Carcinoma Segmentation in Reflectance Confocal Microscopy.

Abstract

Basal cell carcinoma (BCC) is the most frequent subtype of skin cancer, and reflectance confocal microscopy (RCM) provides a non-invasive tool for early lesion assessment. Accurate RCM lesion segmentation remains challenging because expert annotations are limited and multimodal auxiliary information is difficult to exploit effectively. To address this problem, we propose MoST-SAM, a multimodal self-training framework built upon the Segment Anything Model (SAM). The framework leverages both labelled and unlabelled data within a teacher student paradigm, incorporates imaging depth and clinical text as geometric and semantic guidance, and introduces a multi modal confidence assessment mechanism that combines model uncertainty, text-visual consistency, and geometric constraints for hierarchical pseudo-label filtering. We evaluate MoST-SAM primarily on the private MoSKiT-RCM dataset and use the public QaTa-COVID19+ chest X-ray dataset as an auxiliary transfer benchmark. Under low-annotation settings, MoST-SAM improves segmentation performance over representative semi supervised baselines on the evaluated RCM dataset. These results demonstrate that multimodal confidence-guided self-training enables annotation-efficient RCM lesion segmentation and provides a potential foundation for computer-aided analysis of RCM images under limited-annotation settings.