Related Experiment Video
Updated: Sep 7, 2026
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Investigation of S-Detect AI (a computer-aided diagnosis tool) versus ultrasound operators for thyroid nodule
Kieran Walker1, James Michael2, Mark Ballard1
1University Hospitals Birmingham NHS Foundation Trust, UK.
Introduction:
Thyroid nodules affect more than 60% of the population, approximately 5% being malignant. The British Thyroid Association U-classification guides ultrasound risk stratification across the United Kingdom but is subject to inter-observer variability. Artificial intelligence decision support tools may offer more standardised assessment, yet no peer-reviewed study has validated S-Detect, a computer-aided diagnosis tool integrated into Samsung ultrasound systems, against the British Thyroid Association framework. This study compares S-Detect with experienced ultrasound operators using histology as the reference standard.
Methods:
This retrospective pilot diagnostic accuracy study assessed 32 thyroid nodules (September 2023 to August 2024). S-Detect served as the index test applied retrospectively on a Samsung RS85 Prestige platform; prospectively assigned U-scores were the comparator and surgical histology the reference standard. A secondary analysis varied Doppler input across all four options in the 11 malignancies misclassified by S-Detect to test whether Doppler selection contributed to misclassification.
Results:
Operators achieved superior balanced performance (sensitivity: 78.9%, specificity: 84.6%, accuracy: 81.3%, and area under the curve: 0.870) compared with S-Detect (42.1%, 100%, 65.6%, area under the curve: 0.729), which under-classified 58% of malignancies. Inter-method agreement was slight (κ = 0.175), with significant directional discordance (McNemar p = 0.008). Doppler manipulation produced substantial variability: 81.8% of misclassified cases showed U-score change attributable solely to Doppler selection, with peripheral patterns yielding zero correct malignancy classifications.
Conclusion:
Operators achieved superior diagnostic performance. S-Detect's high false-negative rate appears largely driven by over-reliance on Doppler. Its clinical utility is limited within this single-centre study. Future iterations excluding Doppler would likely achieve substantially improved sensitivity.
