A similarity-aware network with contrastive optimization for biomedical image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate biomedical image segmentation is vital for clinical diagnosis.
• Deep learning models (CNNs, Transformers) excel but require large annotated datasets, which are scarce.
• Existing methods often underutilize geometric consistency from unlabeled data.

Purpose of the Study:

• To develop an annotation-efficient method for biomedical image segmentation.
• To improve segmentation accuracy by incorporating global shape consistency.
• To leverage unlabeled data effectively without increasing computational cost.

Main Methods:

• Proposed SimBIS (similarity-aware network with contrastive optimization).
• Integrated output-space consistency using Hu-moments to capture global mask geometry.
• Minimized discrepancy between segmentation predictions from augmented views in Hu-moment space.
• Retained pixel-wise supervised loss for fine-grained boundary optimization.

Main Results:

• SimBIS improved segmentation performance (mDice, mIoU) on LID and SID datasets.
• Outperformed state-of-the-art methods on Kvasir-Seg, CVC-ClinicDB, and ISIC 2018 datasets.
• Achieved more accurate and consistent contours, effectively segmenting small targets and challenging structures.

Conclusions:

• SimBIS enhances biomedical image segmentation via Hu-moment shape consistency and contrastive learning.
• Effectively leverages augmented and unlabeled data without extra parameters or inference cost.
• Shows strong potential for annotation-efficient biomedical image segmentation.