Border Feature Enhancement in Ultrasound Image Segmentation

Semantic segmentation of ultrasonic images is a crucial and challenging task in computer vision in AI-assisted medicine. In prevailing methods, each pixel in the image is treated equally while ignoring the fact that border pixels contain more critical information. This limitation reduces segmentation performance and may result in irregular and ambiguous segmentation results, even under high-quality manual labels. To address this issue, we propose a novel boundary enhancement scheme that improves segmentation accuracy by incorporating a complementary decoder structure (Bor-decoder). This dual-decoder scheme simultaneously predicts both the target area and its border in ultrasonic images, with feature fusion between these two decoders to boost overall performance. Our approach automatically generates border labels from segmentation mask annotations without requiring additional manual border annotations. Experimental results show that the two baseline models enhanced with our boundary enhancement scheme outperform state-of-the-art methods, achieving the highest Dice scores on three datasets: TN3k 86.41% (85.45% without Bor-decoder), DDTI 82.97% (82.54%), and FLUIS 83.82% (82.43%). Furthermore, our method significantly improves the detection of small and multiple targets in ultrasound images, enhancing clinical diagnostic accuracy and supporting more reliable medical assessment.