An attention and denoising enhanced network for semi-supervised left atrial segmentation

Accurate left atrial (LA) segmentation is vital for cardiovascular disease management. However, current meth- ods struggle with complex LA structures, noise, and especially limited annotated data. This paper introduces ADeNet, a novel semi-supervised framework (Attention and Denoising Enhancement Network) that effectively leverages unlabeled data to improve segmentation accuracy and robustness. ADeNet integrates an enhanced multi-scale self-attention mechanism, an adaptive denoising module, and a multi-decoder consistency regular- ization strategy. By incorporating multi-scale self-attention into an encoder-decoder architecture with diverse upsampling decoders, ADeNet better adapts to complex LA shapes and captures both global and local contex- tual information. The adaptive denoising module estimates uncertainty from decoder discrepancies and dynam- ically adjusts loss weights, improving noise robustness. Crucially, the consistency regularization scheme enforces consistent predictions across multiple decoders under diverse perturbations, including data augmentation and strategically applied feature-level noise. This maximizes the information extracted from unlabeled data. Exper- iments on a public LA dataset show ADeNet significantly outperforms state-ofthe-art methods in accuracy and robustness, offering an effective approach for high-precision LA segmentation.