Toward Improved Vault Prediction: Quantitative Annotation Of Anterior Ciliary Body Morphology On Ubm Using Labelme With Deep Learning Validation

Anterior ciliary body morphology is an important anatomical factor influencing postoperative vault after implantable lens surgery. Ultrasound biomicroscopy (UBM) allows detailed visualization of these structures; however, standardized quantitative annotation methods and objective validation strategies remain limited. This study aimed to establish a Labelme-based annotation protocol for anterior ciliary body morphology on UBM images and to evaluate its reliability using statistical, morphological, and deep learning–based validation.

Single-center retrospective imaging study conducted at a university-affiliated tertiary eye care center.

This retrospective study included UBM images from 99 subjects (20 images per subject). A trained grader annotated key anatomical structures using Labelme, including the posterior iris root surface, anterior ciliary body recess, and radial folds of the anterior ciliary body.

Reliability was assessed through:
(1) classification statistics to evaluate morphological distribution;
(2) morphological verification to confirm consistency with predefined structural categories;
(3) deep learning validation using a YOLO-based detection model trained for 32 epochs, with performance evaluated by box loss, classification loss, DFL loss, precision, recall, mAP50, and mAP50–95.

All images were successfully annotated. Morphological classification showed stable proportional distributions, and verification confirmed consistency between annotations and anatomical definitions.

During training, box loss decreased from 1.426 to 0.652, classification loss from 2.633 to 0.412, and DFL loss from 1.492 to 0.985. Precision and recall rapidly reached 1.000. In validation, box loss declined from 0.904 to 0.556, classification loss from 2.446 to 0.375, and DFL converged to 0.981. Detection performance showed mAP50 increased to approximately 0.995, while mAP50–95 improved from 0.461 to 0.877.

A standardized and reliable protocol for quantitative annotation of anterior ciliary body morphology on UBM images was established. Combined statistical, morphological, and deep learning validation demonstrated high robustness and accuracy, providing a methodological foundation for future vault prediction research in implantable lens surgery.