Deep learning-based image segmentation of anterior segment UBM images for primary angle-closure glaucoma

DOI: 10.3760/cma.j.cn115989-20250513-00151

Published: 2025 -11 -10

Citation

Yu Xinqi, Zhao Zhiyuan, Miao Qinghao, et al. Deep learning-based image segmentation of anterior segment UBM images for primary angle-closure glaucoma[J]. Chin J Exp Ophthalmol, 2025, 43(11):1017-1023. DOI: 10.3760/cma.j.cn115989-20250513-00151.

ABSTRACT [Download PDF] [Read Full Text]

Objective To develop a deep learning-based segmentation model for anterior segment ultrasound biomicroscopy (UBM) images to automatically segment the anterior segment tissues of patients with primary angle-closure glaucoma (PACG).

Methods A single-center retrospective case series was conducted.A small-scale dataset comprised 468 UBM images of the anterior chamber angle closure from 156 patients with PACG who underwent the UBM examination at Tianjin Medical University Eye Hospital between July 12, 2022, and February 20, 2023.The UBM images were randomly split into a training dataset of 228 images and a testing dataset of 152 images using a random seed method in a ratio of 6∶4.The models were trained using the PSPNet model with MobileNet V2 and ResNet50 as backbones, the DeepLab v3+ model with MobileNet V2 and Xception as backbones, and the SegFormer model with MiT-B0 and MiT-B2 as backbones.The testing dataset was used for result prediction and to achieve segmentation of four regions: the cornea and sclera, iris, ciliary body, and anterior lens surface.To evaluate the performance of the models in segmenting the anterior segment structures, multiple metrics were assessed, including the mean intersection over union (mIoU), Dice coefficient, precision, recall, false negative rate, and specificity.A comparative analysis of the test results across the different models was subsequently performed.This study adhered to the Declaration of Helsinki.The study protocol was approved by the Ethics Committee of Tianjin Medical University Eye Hospital (No.2023KY-05).

Results The two models with the best segmentation performance were PSPNet and DeepLab v3+ .The PSPNet model with ResNet50 as the backbone achieved the mIoU of 85.11%, Dice coefficient of 91.38%, precision of 91.83%, recall of 90.94%, false negative rate of 9.06%, and specificity of 98.89%.The DeepLab v3+ model with MobileNet V2 as the backbone achieved an mIoU of 85.84%, Dice coefficient of 92.01%, precision of 92.67%, recall of 91.36%, false negative rate of 8.64%, and specificity of 98.90%.Among the five key metrics, mIoU, Dice coefficient, recall, false negative rate, and specificity, DeepLab v3+ exhibited the best segmentation performance.In addition, the DeepLab v3+ model with Xception as the backbone had the highest precision among all models, reaching 92.77%.

Conclusions The deep learning-based DeepLab v3+ model achieves precise segmentation of anterior segment tissue structures in PACG anterior segment UBM image segmentation, providing auxiliary support for clinical diagnosis.