MFSE-TransUNet: A Thyroid Nodule Ultrasound Image Segmentation Network Integrated With Dynamic Feature Calibration and Edge Enhancement

Ultrasound imaging is a commonly used auxiliary diagnostic method for detecting thyroid nodules. However, its low resolution, high noise interference, numerous artifacts, and blurred boundaries make manual annotation time-consuming and highly subjective. Thus, accurate pixel-level segmentation is of great value for quantifying nodule morphology, tracking lesion progression, and planning surgery. Although the existing TransUNet balances local and global features through a hybrid CNN-Transformer architecture, it still faces the following challenges in thyroid nodule segmentation: firstly, the fixed convolutional kernels struggle to adapt to nodule morphological diversity; Secondly, existing multi-scale feature fusion methods fail to consider hierarchical contribution differences; Furthermore, significant edge information is easily lost during upsampling. Accordingly, this study proposes the MFSE-TransUNet model to address these issues. Experimental results on the TUD, TN3K, and DDTI datasets demonstrate the model’s effectiveness, achieving MIoU improvements of 10.89%, 6.02%, and 7.83%, and Dice coefficient increases of 7.79%, 3.79%, and 5.21% compared to the TransUNet baseline. All metrics show consistent improvements, with cross-dataset Dice fluctuations of less than 1.5%, strongly demonstrating the model’s generalization capability.