GraphCL: Graph-based Clustering for Semi-Supervised Medical Image Segmentation

TL;DR

This work proposes a graph-based clustering for semi-supervised medical image segmentation (GraphCL) by jointly modeling graph data structure in a unified deep model, and is the first work to model the data structure information for semi-supervised medical image segmentation (SSMIS).

Abstract

Semi-supervised learning (SSL) has made notable advancements in medical image segmentation (MIS), particularly in scenarios with limited labeled data and significantly enhancing data utilization efficiency. Previous methods primarily focus on complex training strategies to utilize unlabeled data but neglect the importance of graph structural information. Different from existing methods, we propose a graph-based clustering for semi-supervised medical image segmentation (GraphCL) by jointly modeling graph data structure in a unified deep model. The proposed GraphCL model enjoys several advantages. Firstly, to the best of our knowledge, this is the first work to model the data structure information for semi-supervised medical image segmentation (SSMIS). Secondly, to get the clustered features across different graphs, we integrate both pairwise affinities between local image features and raw features as inputs. Extensive experimental results on three standard benchmarks show that the proposed GraphCL algorithm outperforms state-of-the-art semi-supervised medical image segmentation methods.

Figures (6)

• Figure 1: We apply graph neural networks (GNN) to address SSMIS challenges. Specifically, we create a graph representation to capture the data structure, and then use GNN-based clustering to group the graphs.
• Figure 2: The proposed Graph-based Clustering for Semi-supervised Medical Image Segmentation (GraphCL) architecture consists of two core graph mechanisms: GCN alignment and clustering graph construction. In the GCN alignment phase, a data structure analysis network generates structured scores containing structural information, while the CNN is responsible for feature extraction. These structured scores are combined with the CNN-extracted features to construct a dense instance graph for the GCN. After merging the features from both CNN and GCN, the system inputs them to align data from the same category. Regarding the clustering graph construction, we create a similarity matrix based on the similarity between local features, which is then used as the adjacency matrix for the graph. Finally, we utilize this adjacency matrix and deep features as node features to complete the graph construction.
• Figure 3: Sensitivity analysis on the ACDC dataset with the labeled ratio of 3(5%) .
• Figure 4: Sensitivity analysis on the ACDC dataset with the labeled ratio of 7(10%) .
• Figure 5: Kernel dense estimations of different methods, trained on 5% labeled ACDC dataset. Top to bottom are kernel-dense estimations of features belonging to three different classes of ACDC: left ventricle, myocardium, and right ventricle.