Advancement of cataract classification with artificial intelligence using anterior segment optical coherence...

To evaluate the efficacy of a self-supervised learning Vision Transformer (ViT) for classification of the nucleus, cortex, and posterior capsule cataract severity utilizing anterior segment optical coherence tomography (AS-OCT) images.

Artificial intelligence (AI) model training.

Overall, 1,693 eyes were imaged using AS-OCT, with 1,023 classified according to the Lens Opacities Classification System III for supervised training at the Department of Ophthalmology, University of Tsukuba Hospital, Japan. Five AI models were compared: ResNet18, ViT with/without ImageNet pre-training, and Self-Supervised ViT (SS-ViT) constructed using AS-OCT images from 670 eyes. These models were evaluated across five classification tasks: nuclear cataract 2-class (N1 vs. N2≥), 3-class (N1, N2-N3, N4≥), and 6-class classifications; and 2-class classifications for both cortical (C1 vs. C2≥) and posterior subcapsular cataracts (P1 vs. P2≥). Performance was measured using Area Under the Precision-Recall Curve (AUPRC).

In the nuclear cataract 2-class classification, ResNet18, ViT, and SS-ViT, which were pre-trained on ImageNet, demonstrated the highest AUPRC of 0.999. For the nuclear cataract 3-class classification, SS-ViT exhibited the highest AUPRC of 0.939. In the 6-class classification of nuclear cataract, SS-ViT attained the highest value with an AUPRC of 0.788. In the 2-class classification of cortical cataract, SS-ViT demonstrated the highest performance with an AUPRC of 0.774, while in the 2-class classification of posterior subcapsular cataract, SS-ViT exhibited the most favorable performance with an AUPRC of 0.506.

Our findings demonstrate the effectiveness of a self-supervised ViT model for severity classification of nuclear, cortical, and posterior subcapsular cataracts on AS-OCT.