Standardized evaluation of orthokeratology lens fitting status assisted by deep learning algorithm.

Abstract

Objective: Aims to develop a standardized evaluation method based on the YOLOv8n object detection algorithm, with the goal of quantitatively assessing the displacement of orthokeratology lenses (OK_lenses) and objectively evaluating their fitting status under fluorescein staining conditions.

Methods: 117 videos documenting of the fluorescein sodium staining fitting process of OK_lenses, were analyzed alongside corresponding ophthalmic examination results. Key frames were extracted from these videos, annotated, and compiled into a local dataset. A YOLOv8n object detection algorithm model was constructed to automatically identify the visible iris and OK_lenses. The lens displacement was calculated based on the distance between the center points of these two targets and the horizontal visible iris diameter (HVID). Model performance was evaluated using Accuracy, Recall, and F1-score metrics, while consistency was assessed via the Kappa consistency test and compared with traditional manual assessment results.

Results: Achieved automatic calculation of OK_lens displacement in fluorescein-stained images and analysis of their fitting status through the application of the YOLOv8n object detection algorithm. For iris and lens detection, the model yielded accuracy, recall, and F1-score values of 97.6 %, 98.8 %, 98.2 % and 98.4 %, 98.9 %, 98.6 %, respectively. In terms of fitting status assessment, the accuracy of the three prediction results of this model compared with the gold standard was 88.6 %, 91.4 % and 88.6 % respectively, the average accuracy rate was 89.5 %, and all Kappa values were > 0.80, showing a high consistency with the gold standard.

Conclusion: The YOLOv8n object detection algorithm model demonstrates strong concordance with the gold standard and high accuracy, validating its robust replication of the gold standard's evaluation logic. By enabling the objective, quantitative analysis of lens displacement in fluorescein-stained images, this highly stable and objective model could serve as a standardized tool to complement traditional manual evaluations. It offers considerable value for improving assessment efficiency of physicians lacking fitting experience and reducing subjective bias during fitting procedures. Notably, this study may not improve efficiency for experienced fitting physicians, as the steps and time required for fitting may not be significantly reduced. To further verify its generalization ability, future studies should expand the sample size and incorporate multi-center data for comprehensive validation to support its transition toward clinical standardization.