A deep learning-based method for evaluating the fitting states of orthokeratology lenses using fluorescein staining videos
Related Experiment Video
Updated: Mar 9, 2026
Spatio-Temporal In Vivo Imaging of Ocular Drug Delivery Systems using Fiberoptic Confocal Laser Microendoscopy
Published on: September 27, 2021
Rong Lu1, Zhiqiang He1, Kai Niu1
1Key Laboratory of Universal Wireless Communications, Beijing University of Posts and Telecommunications, Beijing, 100876, China.
Purpose:
To develop and validate an automated deep learning-based tool for evaluating the fitting states of orthokeratology lenses using a dual-stream architecture based on ResNet-50 and a temporal attention module, compare its performance to the manual annotation results of experienced optometrists and assess the effectiveness of the methods used in the model.
Methods:
The dataset comprises 143 video recordings of fluorescein staining examinations conducted under slit-lamp microscopy. Five experienced optometrists independently evaluated these videos, classifying lens fitting states into three categories: loose-fitting, well-fitting, and tight-fitting. The consensus assessment of these optometrists was adopted as the gold standard. The ResNet-50 network was utilized to process intra-frame images and extract lens morphological features, while a temporal attention module was employed to analyze successive inter-frame images and capture lens movement features. To enhance the capture of spatial characteristics of lens fitting, an edge detection method was used as a form of data augmentation. After independent feature extraction, the outputs from the two branches were aggregated through concatenation, with the fused result being used for the final classification task.
Results:
By incorporating annotations from experienced optometrists, the proposed model demonstrates strong performance in classifying orthokeratology (Ortho-K) lens fitting states, achieving an overall accuracy of 92.3%, a macro sensitivity of 92.6%, and a macro specificity of 96.2%. The approach outperforms individual TimeSformer and ViViT models, with the integration of the proposed module into these models resulting in a significant enhancement in evaluation accuracy.
Conclusion:
The study demonstrated the effectiveness of the proposed model, which could automatically and accurately evaluate the fitting states of orthokeratology lenses. This approach offers a reliable and objective evaluation method, significantly aiding in the clinical assessment of orthokeratology lens fitting.
