Edge Harmony Attention Network for Semi-Supervised Medical Image Segmentation

In the Medical field, image segmentation plays a crucial role in the identification and quantification of lesions, tumors, and other abnormal structures. However, obtaining a substantial quantity of well-annotated, high-quality images remains a significant challenge due to the high cost and labor-intensive nature of annotation. To address this issue, some studies have attempted to augment datasets by generating synthetic images, but discontinuities at the boundaries of synthetic images can negatively impact segmentation accuracy. In this work, we propose the Edge Harmony Attention Network (EHA-Net), a novel semi-supervised model for medical image segmentation. The proposed model incorporates an Edge Harmonization Module that mitigates the edge artifacts commonly associated with copy-paste augmentation techniques, ensuring smooth and natural transitions along image boundaries, thereby improving segmentation accuracy. Additionally, we introduce a Foreground Attention Module to enhance feature alignment between the foreground and background, promoting spatial consistency and focusing on coherent image details. Experimental results demonstrate that EHA-Net significantly outperforms existing state-of-the-art semi-supervised approaches across four publicly available medical imaging datasets. The code is accessible on GitHub at https://github. com/Manty1998/EHA Net.