Edge guidance unetr for malignant tumor segmentation in breast ultrasound image

Breast ultrasound (BUS) images are pivotal for early tumor detection, yet accurate segmentation faces challenges like noise and indistinct boundaries. Existing methods, such as multi-task learning and adaptive attention networks, often overlook morphological features intrinsic to malignant tumors, resulting in suboptimal performance. In response, our study introduces EGUnetr, a novel segmentation model with Edge Guidance strategy designed to address these limitations. EGUnetr incorporates an improved Transformer block using the Local-Global shifted window (LGWT), enhancing the encoder's ability to learn robust representations of malignant tumor samples in breast ultrasound images. Additionally, we develop an Edge Guidance Module to extract and fuse target boundary features with semantically rich information. This guides the decoder to generate more precise prediction maps, offering a detailed and accurate portrayal of breast tumor boundaries. Experimental results demonstrate the superior performance of EGUnetr in malignant tumor segmentation over existing medical image segmentation models.