Asymmetric Network Based on CNN and Attention Mechanisms for Thyroid Nodule Segmentation

Thyroid nodules are detected in up to 65% of the general population, and their early identification is clinically significant for formulating personalized treatment plans. Precise segmentation of nodules in ultrasound images is essential for assessing their morphological and pathological characteristics. In current clinical practice, the annotation of nodule regions primarily relies on manual operations performed by physicians. This process is not only time-consuming but also labor-intensive. Furthermore, due to variations in physicians’ professional experience, the annotation results are difficult to ensure consistency and reliability. This study proposes an asymmetric encoder-decoder segmentation architecture based on Convolutional Neural Networks (CNN) and attention mechanisms for automatically and precisely segmenting thyroid nodules in ultrasound images. The framework introduces an Efficient Convolutional Block (ECB) to extract high-level semantic features, constructs a Convolutional Modulation Module (CMM) to enhance feature representation, and incorporates a Spatial Semantic Enhancement Module (SSEM) to optimize detail reconstruction. Experimental results on a public thyroid nodule ultrasound dataset demonstrate that the proposed method significantly outperforms existing models’ overall segmentation performance, achieving a Dice Similarity Coefficient (DSC) of 0.8332 and an Intersection over Union (IOU) of 0.7473 on the test set. Ablation studies and visual analysis further confirm the effectiveness and interpretability of each module.