Abstract: To address the issues of noise interference,low contrast,and blurred boundaries in thyroid nodule ultrasound images,an improved SAMUS  was constructed based on the SAMUS framework.This model employs a shallow feature extraction module to obtain multi-scale detail information and introduces a shallow feature fusion strategy in the decoding stage to supplement shallow representations and enhance adaptability to morphological changes of nodules.A comparative experiment conducted between the improved SAMUS and four traditional models shows that compared with the U-Net,U-Net v2,DAEFormer and SAMUS,the improved SAMUS  can more accurately restore the target shape,effectively reduce mis-segmentation and under-segmentation phenomena;compared with the SAMUS,the IOU of the improved SAMUS  increases by 0.89,the Dice increases by 0.73,and the HD95 decreases by 0.30.This indicates that the improved SAMUS can obtain shallow multi-scale detailed features and fuse them during the decoding stage,which can improve the overlap degree of thyroid nodule image segmentation and reduce boundary errors.

Key words: thyroid nodules, ultrasound image segmentation, multi-scale fusion, shallow feature extraction, SAMUS, deep learning