Intra-Class Subdivision for Pixel Contrastive Learning: Application to Semi-supervised Cardiac Image Segmentation

TL;DR

This work addresses boundary-induced representation contamination in semi-supervised cardiac image segmentation by partitioning each class into inner and boundary subclasses and introducing the unconcerned sample to reduce intra-class variance. It proposes SPCL, a teacher-student framework that applies separate inner and boundary contrastive losses (ICL and BCL) with a reliable-aware sampling strategy and memory banks to promote compact intra-subclass features and improved inter-class separation. The method achieves state-of-the-art or competitive results on SCD, ACDC, and M&Ms under limited supervision, with ablations confirming the value of unconcerned samples and boundary-focused learning. Overall, SPCL demonstrates that modeling fine-grained subclass structure and boundary discrimination yields better segmentation quality and boundary precision in medical imaging at low annotation budgets.

Abstract

We propose an intra-class subdivision pixel contrastive learning (SPCL) framework for cardiac image segmentation to address representation contamination at boundaries. The novel concept ``Unconcerned sample'' is proposed to distinguish pixel representations at the inner and boundary regions within the same class, facilitating a clearer characterization of intra-class variations. A novel boundary contrastive loss for boundary representations is proposed to enhance representation discrimination across boundaries. The advantages of the unconcerned sample and boundary contrastive loss are analyzed theoretically. Experimental results in public cardiac datasets demonstrate that SPCL significantly improves segmentation performance, outperforming existing methods with respect to segmentation quality and boundary precision. Our code is available at https://github.com/Jrstud203/SPCL.

Figures (8)

• Figure 1: Comparison of existing BACL versus our SPCL. BACL methods pull both inner and boundary samples toward the same centroid, while our SPCL aligns them to separate centroids, addressing intra-class variations and enabling both to act as negatives in inter-class contrastive learning.
• Figure 2: Overview of our SPCL method built on a teacher-student network. Based on the pseudo labels of unlabeled images and the ground truth of labeled images, the representations within the feature maps are categorized into inner and boundary subclasses. Subsequently, inner and boundary contrastive learning are applied to learn the representations of the inner and boundary regions, respectively.
• Figure 3: Influence of unconcerned samples (US) on ICL.
• Figure 4: Visualization of feature space obtained by different BACL methods.
• Figure 4: Ablation study for subclass relation.