A Deep Learning-Based Study on Automatic Measurement Method of Biological Parameters in Anterior Segment UBM Images...

This study aimed to develop a deep learning-based model for measuring biometric parameters from anterior segment ultrasound biomicroscopy (UBM) images, assisting clinicians in the early screening and diagnosis of primary angle-closure glaucoma (PACG).

Through comparative selection, the best-performing model was used to segment four regions in UBM images of PACG patients: cornea and sclera, iris, ciliary body, and lens anterior surface. Using YOLOv11s, the model performed localization of four key points on both pre- and post-segmentation datasets and automatically measured seven biometric parameters related to the angle closure mechanism.

The DeepLabv3+ model performed excellently in the segmentation tasks, with an mean intersection over union (mIoU) of 85.84%, precision of 92.67%, recall of 91.36%, and Dice coefficient of 92.01%, all representing optimal values. In the object detection task, the segmented dataset achieved a precision of 88.3%, recall of 89.9%, and mean average precision at 0.50 IoU (mAP50) of 92.9%, showing at least a 19% improvement over the original image dataset. The average absolute error of the Euclidean distance for four-point localization was 50.41µm, with a root mean square error of 77.85 µm. In biometric parameter measurements, except for the iris-lens angle, all other parameters exhibited intraclass correlation coefficient (ICC) values greater than 0.94.

This study showed that the proposed deep learning-based method for automatic measurement of closure mechanism-related biometric parameters is accurate and effective.

This is the first study, to our knowledge, to integrate multi-region segmentation and four-point localization for the automatic computation of seven anterior segment parameters directly from UBM images, providing precise and reliable data support for early PACG diagnosis and pathogenesis research.