Development and Validation of Artificial Intelligence-driven Cataract Detection Model Using Smartphone-captured Images

Abstract

Purpose : Cataracts are one of the leading causes of blindness worldwide, particularly in resource-limited regions with minimal resources to diagnose or intervene in a timely manner. Inexpensive, effective screening methods – such as cataract detection using commonplace smartphone cameras – may have strong implications for improving accessibility and visual outcomes for the global population. This study was conducted to evaluate the feasibility of a smartphone-based artificial intelligence (AI) algorithm for detecting and grading cataracts.

Methods : This cross-sectional study was conducted between April 2020 and March 2022. Images of cataracts were captured using smartphones (iPhone and Xperia) and slit-lamp microscopy. Three ophthalmologists graded the images based on Emery-Little (EL) classification. Concordance rate between AI and Ophthalmologists was also evaluated in this study. A cataract grade estimation model was developed using the “Neural Network Libraries” software, which stratified the images into two classes: A) ≤grade 2 or B) ≥grade 3. Diagnostic performance of the software was evaluated using sensitivity and specificity for determining cataracts above grade 3.

Results : 1,821 iPhone and 1,889 Xperia images of the cataracts were collected from 178 participants. A cataract grade estimation model was created using 2,548 smartphone images (EL Grade 1, 768; Grade 2, 1,521; and Grade 3≥, 259 images). Concordance rates yielded by the cataract grade estimation model for 1,162 iPhone and Xperia images were 78.7% and 87.2% for Xperia on dilated eyes and 89.5% and 81.5% on non-dilated eyes, respectively. For non-dilated eyes, sensitivities/specificities from iPhone and Xperia were 62.2% / 93.6% and 38.4% / 96.3%, respectively. For dilated eyes, sensitivities/specificities from iPhone and Xperia were 92.3% / 76.6% and 83.3% / 87.9%, respectively.

Conclusions : Machine learning and AI analysis of raw images from smartphone cameras appears to have potential in detecting moderate-to-severe nuclear cataracts. In resource poor or rural settings without traditional tools to visualize the lens (i.e., slit-lamp microscope), this may have implications in promoting early detection and intervention to reduce global prevalence of visual impairment/blindness, as well as to minimize surgical complications associated with delayed treatment.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.