Purpose: To examine the effectiveness of automated pterygium severity grading by utilizing segmentation of photopic anterior segment images from slit-lamp microscopes.
Subjects and methods: Datasets included 438 anterior segment images from 366 cases in Tsukazaki Hospital for training and evaluation, and 142 anterior segment images from 142 cases in Miyata Eye Hospital for testing. There were half normal and half pterygium cases in training and evaluation datasets. Areas of the cornea and pterygium were annotated, and a model for detecting these areas was built using U-Net and DenseNet-121. The center and diameter of the cornea and the edge of the pterygium were obtained from segmented results, and the pterygium severity was measured automatically using the criteria proposed by Miyata et al. (Grade from 0 to 3). Accuracy rates between gradings from annotated images and automated measurements were compared. Accuracy rates for grading when making a clinical diagnosis by physicians were also evaluated.
Results: The accuracy rate for grading in automated measurement was 92.3% (131/142), and 4 (2.8%) and 7 (4.9%) were over- and under-graded, respectively. The agreed accuracy rate for grading by physicians was 69.0% (98/142), and 20 (14.1%) and 24 (16.9%) were over- and under-graded, respectively.
Conclusion: Pterygium severity could be graded with the automated measurement utilizing segmentation machine learning, suggesting its effectiveness in clinical practice.
Nippon Ganka Gakkai Zasshi (J Jpn Ophthalmol Soc) 128: 416-420, 2024.