Development and validation of an Automated computed tomography segmentation model for thyroid nodules using a channel attention high-resolution network.

Background: Thyroid nodules are highly prevalent. Accurate segmentation of thyroid nodules is a prerequisite for computed tomography (CT)-based radiomics and computer-aided diagnosis, but manual region-of-interest delineation is time-consuming, subjective, and difficult to standardize. Although ultrasound remains the first-line imaging modality for thyroid nodule assessment, contrast-enhanced CT is frequently used in selected patients who have complex anatomy, suspected local extension, or require preoperative evaluation. Current deep learning models suffer from insufficient boundary capture for irregular nodules and high computational complexity, which has led to a lack of clinically practical tools. This study aimed to develop and internally validate an efficient automated CT segmentation model for thyroid nodules. Methods: This single-center retrospective study included CT images from 500 patients with pathologically confirmed thyroid nodules (250 benign and 250 malignant). Patients were divided at the patient level into training, validation, and internal test sets comprising 350, 50, and 100 patients, respectively. Eligible patients had thyroid CT and histopathological confirmation, and manual lesion contours served as the ground truth for segmentation. We propose Channel Attention High-Resolution Network (CA-HRNet), which enhances High-Resolution Network (HRNet) with a Channel Feature Selection Module (CFSM) that dynamically fuses multi-scale features by retaining discriminative channels and pruning redundant features, balancing representational capacity and computational efficiency. The model was trained using Dice loss with the RAdam optimizer and data augmentation. Segmentation performance was evaluated using the Dice coefficient (DC) and intersection over union (IoU), and CA-HRNet was compared with U-Net, SegFormer, Transformer U-Net​(TransUNet), DAC-Net, and HRNet. Test-time augmentation (TTA) was applied during inference. Results: The cohort had a median age of 48 years, 373 female patients (74.6%), and a median nodule diameter of 1.60 cm. On the internal test set, CA-HRNet + TTA achieved the best overall segmentation performance, with a DC of 78.6% and an IoU of 70.0%. Performance was higher for benign nodules (Dice, 85.5%) than for malignant nodules (Dice, 66.2%), reflecting the more irregular morphology and less distinct boundaries of malignant lesions. Ablation experiments showed that the combination of CFSM and Channel Attention Convolution Module (CACM) achieved the best accuracy-efficiency balance and reduced computational complexity by 73%, from 93.764 to 25.376 giga multiply-accumulate operations (Giga MACs). Qualitative assessment showed that CA-HRNet produced masks with improved boundary adherence and detail preservation compared with competing models. Conclusions: CA-HRNet achieved accurate and computationally efficient automated segmentation of thyroid nodules on a single-center internal CT test set. The model may support reproducible region of interest (ROI) generation for CT-based radiomics and computer-aided diagnosis; however, multicenter external validation and downstream diagnostic or prognostic testing are required before routine clinical implementation.