A deep learning-based classification method for subclinical zonular laxity in AS-OCT images.

Objective: In this study, we developed and validated a deep learning method for the detection and angular position identification of subclinical zonular laxity using anterior segment optical coherence tomography (AS-OCT). Methods: A total of 600 curated AS-OCT images from 536 patients (600 images) undergoing cataract surgery were evenly stratified into subclinical zonular laxity (n = 300 images from 297 patients) and normal control (n = 300 images from 239 patients) groups. Data were partitioned at the patient level to prevent data leakage, with 60% for training, 15% for validation, and 25% for testing. An additional five clinical cases were used for external validation. We implemented MDCL-Net, a novel classification framework integrating mask-aware feature enhancement, dynamic contextual feature aggregation. Results: The model achieved an accuracy of 79.72%, an area under the receiver operating characteristic curve (AUC) of 86.41%, and an F1-score of 78.93%. Ablation studies confirmed the contribution of each module, and in clinical validation, model-predicted zonular laxity ranges showed good agreement with intraoperative observations across five representative cases. Conclusion: This work presents the first deep learning method capable of both detecting and spatially localizing subclinical zonular abnormalities in AS-OCT images, demonstrating strong clinical applicability and potential as a reliable preoperative screening tool to enhance surgical planning and safety in cataract procedures.