Pairwise Similarity Regularization for Semi-supervised Graph Medical Image Segmentation

TL;DR

This work tackles the distribution shift between labeled and unlabeled data in semi-supervised medical image segmentation by introducing PaSR, a graph-based feature alignment method. PaSR constructs voxel-level graphs from features via a pairwise similarity matrix $\mathcal{A}$, applies Graph Convolutional Networks to obtain refined representations, and jointly optimizes a pairwise similarity loss $d(\mathcal{A}_u,\mathcal{A}_m)$ with a clustering loss to guide pseudo-labeling in a teacher–student framework using bidirectional copy-paste. The approach yields consistent improvements on PROMISE12, ACDC, and Pancreas-NIH, with notably large gains on ACDC at low labeling rates (e.g., average gains exceeding $10\%$). Overall, PaSR demonstrates that aligning cross-domain and intra-domain graph structures, coupled with clustering-informed pseudo-labeling, can enhance robustness and accuracy in semi-supervised medical image segmentation.

Abstract

With fully leveraging the value of unlabeled data, semi-supervised medical image segmentation algorithms significantly reduces the limitation of limited labeled data, achieving a significant improvement in accuracy. However, the distributional shift between labeled and unlabeled data weakens the utilization of information from the labeled data. To alleviate the problem, we propose a graph network feature alignment method based on pairwise similarity regularization (PaSR) for semi-supervised medical image segmentation. PaSR aligns the graph structure of images in different domains by maintaining consistency in the pairwise structural similarity of feature graphs between the target domain and the source domain, reducing distribution shift issues in medical images. Meanwhile, further improving the accuracy of pseudo-labels in the teacher network by aligning graph clustering information to enhance the semi-supervised efficiency of the model. The experimental part was verified on three medical image segmentation benchmark datasets, with results showing improvements over advanced methods in various metrics. On the ACDC dataset, it achieved an average improvement of more than 10.66%.

Figures (5)

• Figure 1: The left figure shows the overall framework of the model. A graph structure is constructed with each voxel feature of the medical images that have been copied and pasted bidirectionally as nodes. The similarity between node features is calculated as the pairwise similarity matrix of the graph, which serves as the weight of the edges. Then, the aligned pairwise similarity matrix is used as the adjacency matrix, inputted into the GCN for information aggregation. Finally the graph clustering information is used to assist in the segmentation task. The right figure shows the process of pairwise similarity regularization alignment. By reducing the pairwise similarity distance between the teacher-student networks, it aligns the intra-domain and inter-domain similarities of different domain features. Thus guiding the network to generate a more consistent graph structure and transfering the aligned image knowledge to pseudo-label generation.
• Figure 2: PaSR ($\alpha$=0.05) and model variant ($\alpha$=0) in segmentation performance on the ACDC dataset with 5% labeled data.
• Figure 3: Sensitivity analysis on the ACDC dataset.
• Figure 4: Kernel dense estimations of different methods, trained on 10% labeled ACDC dataset. From top to bottom are the three classes in the ACDC: right ventricle, myocardium and left ventricle. The green and blue line represent the labeled and unlabeled data.
• Figure 5: Visualizations of several semi-supervised segmentation methods with 5% labeled data and ground truth on ACDC dataset. The green and red lines represent the ground truth and prediction results.