Related Experiment Video
Updated: Jul 15, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Development and validation of an Automated computed tomography segmentation model for thyroid nodules using a channel
Ning Li1,2, Mingjie Jiang3, Yang Huang4
1Jinhua Graduate Joint Training Base, Zhejiang Chinese Medical University, Jinhua, China.
Gland Surgery
|July 14, 2026
Summary
A new deep learning model, CA-HRNet, efficiently segments thyroid nodules on CT scans, improving accuracy for radiomics and computer-aided diagnosis. This automated tool offers better boundary capture and reduced complexity compared to existing methods.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Thyroid nodules are common, and accurate segmentation is crucial for CT-based radiomics and computer-aided diagnosis.
- Manual segmentation is time-consuming, subjective, and lacks standardization.
- Current deep learning models struggle with irregular nodule boundaries and computational demands, limiting clinical utility.
Purpose of the Study:
- To develop and internally validate an efficient automated CT segmentation model for thyroid nodules.
- To improve the accuracy and efficiency of thyroid nodule segmentation for clinical applications.
- To address limitations of existing deep learning models in capturing irregular nodule morphology.
Main Methods:
- A retrospective study of 500 patients with pathologically confirmed thyroid nodules using CT images.
- Development of the Channel Attention High-Resolution Network (CA-HRNet) incorporating a Channel Feature Selection Module (CFSM) for multi-scale feature fusion.
- Training with Dice loss and RAdam optimizer, with performance evaluated using Dice coefficient (DC) and intersection over union (IoU), compared against other models like U-Net and SegFormer.
Main Results:
- CA-HRNet + Test-Time Augmentation (TTA) achieved the highest segmentation performance (DC 78.6%, IoU 70.0%) on the internal test set.
- The model demonstrated higher accuracy for benign nodules (Dice 85.5%) than malignant nodules (Dice 66.2%).
- Ablation studies confirmed the accuracy-efficiency balance, reducing computational complexity by 73%.
Conclusions:
- CA-HRNet provides accurate and computationally efficient automated segmentation of thyroid nodules on CT scans.
- The model shows potential for reproducible region of interest (ROI) generation in CT-based radiomics and computer-aided diagnosis.
- Further multicenter external validation and downstream testing are necessary for clinical implementation.
Keywords:
High-Resolution Network (HRNet)Thyroid nodule segmentationautomated diagnosischannel attentioncomputed tomography (CT)
