Clinical Application of an AI-Driven Framework for Accurate Endometrial Thickness Measurement in Transvaginal Ultrasound.
Authors
Affiliations (5)
Affiliations (5)
- Department of Medical Imaging, The First Affiliated Hospital of Jinan University, Guangzhou, China; Department of Ultrasound Medicine, Shenzhen Guangming District People's Hospital, Shenzhen, China.
- National-Regional Key Technology Engineering Laboratory for Medical Ultrasound, School of Biomedical Engineering, Health Science Center, Shenzhen University, Shenzhen, China; Marshall Laboratory of Biomedical Engineering, Shenzhen University, Shenzhen, China; Medical UltraSound Image Computing (MUSIC) Lab, School of Biomedical Engineering, Shenzhen University, Shenzhen, China.
- Department of Ultrasound Medicine, Shenzhen Guangming District People's Hospital, Shenzhen, China.
- Shenzhen RayShape Medical Technology Co., Ltd, Shenzhen, China.
- Department of Medical Imaging, The First Affiliated Hospital of Jinan University, Guangzhou, China. Electronic address: [email protected].
Abstract
This study aims to develop an AI-based framework for automatic endometrial thickness (ET) measurement in transvaginal ultrasound (TVUS) based on a large dataset and evaluate its performance from various clinical perspectives. A dataset of 9850 ultrasound images from 5110 cases at Shenzhen Guangming District People's Hospital (2019-2023) was retrospectively included for training and internal validation. For external validation, 356 images from 300 cases were prospectively collected. All images were acquired and annotated by three sonographers (D1-D2-D3, senior-junior-expert) with varying experience levels. The framework includes a uterine corpus segmentation model, an endometrial segmentation model and a maximum interior tangent circle search algorithm for ET measurement. Model performance was assessed using mean absolute error (MAE), intraclass correlation coefficient (ICC) and percentage of measurements within a 2 mm error range (<±2 mm). Clinical acceptability rate was also evaluated. The model achieved a MAE of 1.05 and <±2 mm of 87.40% in the internal validation. For external validation, the MAE of groups AI-D1 and AI-D2 were 0.89 and 1.01. The ICC and <±2 mm for group AI-D1 were 0.90 and 90.18%, respectively. The clinical acceptability rate for AI measurements was 74.85%, lower than D1 (87.37%) but slightly higher than D2 (72.50%). The model completed measurements in 0.2 s, 30 times faster than less experienced human sonographers. The proposed AI framework demonstrates strong potential for clinical application in automated ET measurement with high accuracy, efficiency and clinical acceptability, remaining comparable to experienced sonographers across various clinical perspective. To facilitate further research and clinical translation, the source code for the proposed framework has been made publicly available at https://github.com/bing4250/Endo_segmentation.git.