Related Experiment Video
Updated: Jun 15, 2025
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Adaptive Dual-Task Deep Learning for Automated Thyroid Cancer Triaging at Screening US
Shao-Hong Wu1, Ming-De Li1, Wen-Juan Tong1
1Department of Medical Ultrasonics, Institute of Diagnostic and Interventional Ultrasound, First Affiliated Hospital of Sun Yat-sen University; Ultrasomics Artificial Intelligence X-Laboratory, MedAI Collaborative Laboratory, Guangzhou 510080, China.
An adaptive dual-task deep learning model (ThyNet-S) enhanced thyroid cancer screening efficiency. This artificial intelligence tool improved diagnostic accuracy and reduced unnecessary procedures, optimizing clinical decision-making for thyroid ultrasound screening.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Medicine
- Oncology
Background:
- Thyroid cancer screening relies on ultrasound (US) interpretation, which can be resource-intensive.
- Accurate detection and classification of thyroid lesions are crucial for effective screening.
- Deep learning models offer potential for improving diagnostic performance and workflow efficiency.
Purpose of the Study:
- To develop and evaluate an adaptive dual-task deep learning model, ThyNet-S, for thyroid US screening.
- To assess ThyNet-S's performance in detecting and classifying thyroid lesions and determining cancer risk.
- To investigate the impact of ThyNet-S on radiologist workload and clinical decision-making.
Main Methods:
- A retrospective multicenter dataset of 35,008 thyroid US images from 23,294 examinations was used.
- ThyNet-S integrated lesion detection (pixel-level) and classification (deep semantic features) for cancer risk assessment.
- Diagnostic performance was compared between ThyNet-S-assisted screening and traditional radiologist-only screening using sensitivity, specificity, accuracy, and AUC.
Main Results:
- ThyNet-S-assisted screening demonstrated a higher area under the receiver operating characteristic curve (AUC) compared to radiologists alone (0.93 vs. 0.91/0.92).
- The model improved sensitivity for junior radiologists and significantly reduced radiologist workload by triaging 60.4% of cases.
- Unnecessary fine needle aspiration rates decreased from 38.7% to 14.9% (ThyNet-S alone) or 11.5% (with TI-RADS).
Conclusions:
- ThyNet-S significantly enhances the efficiency and accuracy of thyroid cancer screening using ultrasound.
- The adaptive dual-task deep learning model optimizes clinical decision-making and reduces unnecessary invasive procedures.
- ThyNet-S represents a valuable tool for improving thyroid US screening workflows and patient management.
