A medical image segmentation method based on adaptive graph sparse algorithm under contrastive learning framework

Abstract

Background

Existing U-shaped models have shown significant potential in medical image segmentation. However, their performance is limited due to the constrained receptive field and lack of global reasoning capability in standard U-shaped structures. This paper aims to develop a module for U-shaped structures to enhance feature discrimination and improve segmentation accuracy.

Methods

This paper proposes a medical image segmentation network based on the U-shaped structure under the contrastive learning framework to achieve accurate segmentation of medical lesion areas. Initially, feature maps are extracted using the encoder of the U-shaped structure and mapped into a two-dimensional graph structure. We then propose a Sparse Dual Graph Mapping (SDGM) method to adaptively sparsify the graph structure, creating multiple sparse graph structures with different node attributes and topologies. Node-level and graph-level contrastive learning are defined using different judgments of positive and negative samples within the graph. Finally, supervised and unsupervised losses are aggregated to enhance the model’s discrimination ability, resulting in the final segmentation mask.

Main Results

Experimental results demonstrate that the proposed SGC module is applicable to various U-shaped networks and outperforms existing techniques on multiple datasets. It achieved 94.02% Dice on the honeycomb lung dataset, 91.78% Dice on the ACDC dataset, and 92.43% Dice on the polyp dataset, all showing state-of-the-art performance.

Significance

The proposed Sparse Graph Contrastive (SGC) module can be applied to any U-shaped structure to enhance its performance. This method maintains high correlation and consistency between automatic segmentation results and expert manual segmentation results. It significantly improves the segmentation performance of lesion areas in medical images, assisting doctors in early screening, accurate diagnosis, and adaptive treatment, with important clinical relevance in medical imaging-assisted diagnosis.