Creative Commons source reading · Data quality and algorithm evaluation
Artificial intelligence-based grading of human cataracts: A feasibility study
Latz C; Rothen F; Licht A; Ponto KA; Mirshahi A
Originally published in PLoS One · 2026-08-13
From the article
Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.
Abstract
AbstractThis feasibility study developed and validated an AI algorithm for automatic classification of human lens opacification (cataracts) using anterior segment optical coherence tomography (OCT). The dataset comprised 1,802 unprocessed OCT images from 901 eyes, with cataract severity graded by slit-lamp examination using the Lens Opacities Classification System (LOCS) III. Nuclear opalescence (NO) was stratified as mild (≤3.0), moderate (3.5–4.0), or severe (≥4.5), while posterior subcapsular cataract (PSC) was classified as mild (<4) or severe (≥4). Nuclear color (NC) and cortical cataract (C) were not evaluated. After balancing, images were randomly divided into three sets: 70% for training, 20% for testing and 10% for validation. A convolutional neural network was trained five times. Patients had a mean age of 72 ± 8.8 years,…
Results
The distribution of LOCS III NO grades in the underlying dataset (n = 1760) was centered on medium severity, with most patients having a cataract of LOCS III NO grade 3–4 (Fig 2a). As this dataset only includes cataracts requiring surgery, LOCS III NO grades of 2.5 and lower are rare. The distribution of LOCS III PSC grades of the dataset (n = 1510) is shown in Fig 2b. Since PSC cataract is by definition centered on the central 3 mm of the visual axis it becomes visually significant at an early stage. This is seen in the high proportion of low PSC grades.
Discussion
The feasibility study results indicate that an AI-based approach, utilizing a deep learning algorithm based on a convolutional neural network (CNN), can automatically classify nuclear cataract lens density based on OCT images, confirming previous findings [5,14]. The CNN-based model developed in this project can classify both nuclear cataract and PSC with a reasonable degree of accuracy and can distinguish between mild and severe cases. In contrast to other approaches, pixel intensity analysis for example [9], this AI model assesses features such as structures and patterns within the crystalline lens. This contributes to a more comprehensive image evaluation, potentially resulting in higher accuracy. In 2008, Wong et al. showed a correlation of r = 0.77 between the NO LOCS III score and the nuclear density evaluation based on AS OCT…
Editorial reading guide
- What data created and tested the model, and was evaluation independent?
- What is the cost of a false positive, false negative or out-of-distribution decision?
