MSFFNet: A Multi-scale Shortcut Feature Fusion Network for Ultrasound Thyroid Nodules Segmentation

The automated segmentation of thyroid nodule ultrasound images has emerged as a pivotal direction in the application of machine learning to medical imaging. However, existing networks often combine CNNs and Vision Transformers to compensate for the inherent limitations of convolutions and leverage the respective strengths of both. While they achieve commendable segmentation performance, these hybrid architectures struggle to accurately perceive target objects in specific contexts due to the intrinsic characteristics of ultrasound images. In this paper, we propose a Multi-scale Shortcut Feature Fusion Network (MSFFNet) for achieving precise segmentation of thyroid nodules in ultrasound images. Our MSFFNet incorporates two essential modules: Multi-scale Residual Feature Fusion Module (MRFF) and Shortcut Channel Fusion Module (SCF). MRFF, employing a multi-level residual structure, efficiently captures both global and local information to discern objects of various sizes. SCF, operating at the channel level, facilitates feature integration to effectively promote knowledge transfer. Experimental demonstrates that on both the Thyroid and BUSI datasets, our MSFFNet has achieved F1 scores of 80.49% and 77.00%, along with IoU scores of 70.59% and 68.68% respectively, surpassing the related state-of-the-art models for medical image segmentation.