Deep Learning-Driven Ciliary Body Segmentation in UBM Imaging: advancing automated measurements for glaucoma analysis

This study introduces an automated deep learning model to improve the efficiency and accuracy of glaucoma screening and diagnosis by enabling precise localization and segmentation of key anatomical structures in ultrasound biomicroscopy (UBM) images. Employing an improved ResUNet architecture, the model localizes the scleral spur apex using a Gaussian heatmap as the region of interest (ROI) and segments the ciliary body and surrounding areas through a multi-scale input strategy to enhance segmentation accuracy and robustness. The model achieved outstanding performance, with an average Euclidean distance of 9.863 µm for localization and segmentation metrics including an IoU of 96.85%, Dice coefficient of 98.34%, Pixel Accuracy (PA) of 97.20%, and F1 score of 97.45%. Anatomical analysis revealed significant differences among glaucoma types: the PACG group showed a significantly smaller average trabecular-ciliary angle (TCA, 52.763°) and a larger ciliary body area within 1,000 µm (CBA1000, 1.064 mm²) compared to the normal and POAG groups. A significant negative correlation between TCA and CBA1000 (p<0.05) was observed, particularly in PACG cases. These findings underscore the model’s potential to enhance UBM image analysis, provide critical insights for personalized glaucoma diagnosis and treatment, and facilitate early intervention in high-risk patients. Future work will focus on expanding datasets and incorporating additional anatomical structures to improve generalization and clinical applicability.