SSCLMix: A self-supervised contrastive learning-based data mixing augmentation method.

Abstract

Due to the privacy and scarcity of medical data, deep learning (DL)-based medical image segmentation methods often face challenges of insufficient training data samples and imbalanced training data. These challenges may lead segmentation models to fail in fully learning critical lesion features, thereby degrading model segmentation performance. In the field of medical image processing, data mixing augmentation based on region dropout and mixing regularization is one effective approach to address this problem. However, existing data mixing enhancement methods face challenges such as loss of image structural information and feature misalignment, resulting in inconsistent quality of mixed samples and further hindering the improvement of segmentation model performance. To solve these problems, we propose a self-supervised contrastive learning-based image mixing method (SSCLMix), which contains two steps. First, it classifies the training samples based on the image structural similarity for subsequent image mixing. Second, it uses dual-encoder contrastive learning and a cross-self-attention mechanism for cross-sample modeling to generate mixed images. Additionally, to enhance the quality of mixed images, we introduce a dual- spatial feature perception residual module (DSFPR), which adopts global structural information perception to reduce the destruction of edge textures and regional information. Experimental results for seven medical image segmentation tasks show that, compared with existing data augmentation methods, our method can generate higher-quality mixed samples, thereby bringing more significant improvements to segmentation model metrics. At the same time, it ranks in the upper-middle range in terms of computational efficiency and has certain practicality.