A Pilot Study on Novel Ptotic Eye Dataset: Automated Prediction of Horizontal Corneal Diameter on Digital Photos of Taiwanese Ptotic Patients by Convolutional Neural Networks (CNNs)

Abstract

Purpose : To present the preliminary results of automated measurements for horizontal corneal diameter (white-to-white distance) by a trained deep learning-based artificial intelligence(AI) model.

Methods : A novel dataset of healthy and ptotic eye images was used to develop a deep learning U-net-based model, which performs automated segmentation of the iris. A total of 100 ptotic eyes from the dataset were used for training, with another 100 ptotic eyes utilized for testing the AI model after adequate training. Iris edges were detected automatically based on the results of AI segmentation, then a circle and its center are simulated and represented mathematically in accordance with partially visible iris edges of ptotic eyes. Therefore, A horizontal corneal diameter, white-to-white distance, was calculated using the radius of the simulated circle for ptotic eyes. The AI predicted white-to-white distance was then compared with an attending physician manual measuring the white-to-white distance on the same 100 testing dataset to provide a comparison.

Results : From the 100 image test results predicted by our trained convolutional neural networks (CNNs) model, the mean of the horizontal corneal diameter in ptotic eyes was 11.99mm (range 10.77mm - 13.61mm, SD 0.514). The results of manual measurement by one attending physician in the same 100 test dataset showed that the mean was 12.21mm (range 11.05mm - 13.43mm, SD 0.486). The manual and AI-predicted measurements were well correlated (Pearson r = 0.66, p < 0.01).

Conclusions : Using our novel ptotic eye dataset, the CNN-based AI model demonstrated its potential to predict the horizontal corneal diameter (white-to-white distance) for ptotic digital photos. Further training on the AI model and careful validation for accuracy both need to be performed and presented in the future.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.