SG-UNet: Hybrid self-guided transformer and U-Net fusion for CT image segmentation

Abstract

In recent years, transformer-based paradigms have made substantial inroads in the domain of CT image segmentation, The Swin Transformer has garnered praise for its strong performance, but it often struggles with capturing fine-grained details, especially in complex tasks like CT image segmentation, where distinguishing subtle differences in key areas is challenging. Additionally, due to its fixed window attention mechanism, Swin Transformer tends to overemphasize local features while overlooking global context, leading to insufficient understanding of critical information and potential loss of important details. To address the limitations of the Swin Transformer, we introduce an innovative U-shaped Hybrid Self-Guided Transformer network (SG-UNet), specifically tailored for CT image segmentation. Our approach refines the self-attention mechanism by integrating hybrid attention with self-guided attention. The hybrid attention mechanism employs adaptive fine-grained global self-attention to capture low-level details and guide token assignment in salient regions, while the self-guided attention dynamically reallocates tokens, prioritizing target regions and reducing attention computation for non-target areas. This synergy enables the model to autonomously refine saliency maps and reassign tokens based on regional importance. To enhance training dynamics, we incorporate a combination of CELoss and BDLoss, which improves training stability, mitigates gradient instability, and accelerates convergence. Additionally, a dynamic learning rate adjustment strategy is employed to optimize the model’s learning process in real-time, ensuring smoother convergence and enhanced performance. Empirical validation on the Synapse and lung datasets demonstrates the superior segmentation performance of the Hybrid Self-Guided Transformer UNet, achieving DSC and HD scores of 82.91 % and 16.46 mm on the Synapse dataset, and 98.13 % and 6.34 mm on the lung dataset, respectively. These results underscore both the effectiveness and the advanced capabilities of our model in segmentation tasks.