Abstract
AI algorithms revolutionized medical image analysis, especially in segmentation. Vision-based medical image segmentation, in particular, is critical for extracting precise anatomical structures and pathological regions from imaging data. Among many vision-based deep learning methods, U-Net-based models with skip-connections have become the dominant architecture for medical image segmentation because of their ability to maintain spatial information throughout the network. However, the semantic gap between low and high-level features poses challenges. This study proposes a Dual Cross-Attention (DCA) scheme to address this. DCA comprises Channel Cross-Attention (CCA) for global channel-wise dependencies and Spatial Cross-Attention (SCA) for spatial dependencies across encoder features. These features are integrated into U-Net's skip-connection scheme. The DCA scheme's performance is tested on CT multi-organ, gland, and polyp segmentation tasks. Evaluation against baseline U-Net and variants shows improved segmentation accuracy using metrics like Dice similarity coefficient (DSC). Qualitative analysis confirms visual improvements. Ablation study assesses DCA components' contributions, revealing their effectiveness in bridging the semantic gap and enhancing segmentation accuracy, with DCA's simplicity and efficiency offering practical benefits.