Artificial intelligence for diagnosing nuclear cataracts based on the Emery-Little classification using anterior segment images

Cataracts constitute the foremost global cause of blindness, significantly impacting public health, particularly in underserved regions lacking specialized ophthalmic care. Artificial intelligence (AI) has shown considerable promise in enhancing diagnostic capabilities for cataracts. In Japan, nuclear cataract grading commonly utilizes the Emery-Little classification, providing a standardized framework for assessing severity. However, AI-driven diagnostic systems based on this specific classification remain insufficiently developed. The objective of this study was to develop and validate an AI algorithm capable of diagnosing nuclear cataracts using the Emery-Little classification, leveraging anterior segment images captured by a smartphone-based portable slit-lamp microscope. This retrospective study analyzed anterior segment video recordings from 2,628 eyes (1,314 patients) collected between July 2020 and March 2022 using Smart Eye Camera (SEC). These videos yielded 194,189 static images. Our AI system employed a two-stage machine learning process: first, identifying diagnostically relevant frames via ConvNeXT, and subsequently grading nuclear sclerosis (NS) severity based on the Emery-Little classification using EfficientNetV2. Annotations were independently provided by three ophthalmologists. Diagnostic performance metrics included accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC). The first-stage model achieved high accuracy (0.963), sensitivity (0.958), specificity (0.977), PPV (0.990), NPV (0.906), and AUC (0.967). The second-stage model demonstrated robust performance with accuracy (0.918), sensitivity (0.941), specificity (0.827), and AUC (0.884). Grad-CAM visualization effectively confirmed accurate lens region identification irrespective of pupil dilation. This AI model may be useful for screening or triage of nuclear cataracts, particularly in remote or underserved settings. However, performance for advanced nuclear sclerosis (NS3 + ) was limited; therefore, the current model is not yet reliable for severe cases and should not be used as a stand-alone basis for surgical referral decisions.