Research Article
Behavioral Analysis of Semantic-Guided Self-Supervised Learning for Low-Labeled Medical Image Segmentation
@INPROCEEDINGS{10.4108/eai.22-5-2026.2365099, author={Jihan Fan}, title={Behavioral Analysis of Semantic-Guided Self-Supervised Learning for Low-Labeled Medical Image Segmentation}, proceedings={Proceedings of the 4th International Conference on Image, Algorithms, and Artificial Intelligence, ICIAAI 2026, 22-24 May 2026, Singapore, Singapore}, publisher={EAI}, proceedings_a={ICIAAI}, year={2026}, month={8}, keywords={Medical image segmentation Self-supervised learning Semantic-guided}, doi={10.4108/eai.22-5-2026.2365099} }- Jihan Fan
Year: 2026
Behavioral Analysis of Semantic-Guided Self-Supervised Learning for Low-Labeled Medical Image Segmentation
ICIAAI
EAI
DOI: 10.4108/eai.22-5-2026.2365099
Abstract
This paper focuses on brain tumor segmentation via the BraTS 2018 dataset, reproducing and analyzing typical self-supervised pre-training methods (e.g., SimMIM) under low-label conditions. It then constructs Teacher-Guided SimMIM (TG-SimMIM) as an analytical tool (not a performance optimizer): using foreground probability maps from limited labeled data, it softly guides mask sampling in pre-training to explore its low-label behaviors. Experiments (varying label proportions/random seeds) show: moderate labeling brings limited but stable self-supervised pre-training gains; extremely low labeling makes teacher-guided semantics introduce cumulative bias, reducing model performance. This study notes mask-level semantic guidance risks in low-label medical segmentation, offering evidence for related methods’ applicable boundaries.
