Anterior Segment Imaging Using Responsible Ai: Frameworks For Lmics

This study aimed to develop and evaluate artificial intelligence (AI) models for automated analysis of anterior segment imaging, focusing on anterior segment optical coherence tomography (AS-OCT) and slit-lamp photographs to detect diseases such as glaucoma angle closure, corneal opacities, and keratitis in low- and middle-income countries (LMICs) settings. The purpose was to create empirically grounded frameworks for responsible AI deployment, addressing biases, fairness, and contextual ethics in cross-border healthcare applications like those in India.

The research was conducted using datasets from multicenter studies, including 557,468 AS-OCT scans from 20,542 patients in London and additional images from South India eye hospitals involving 174 patients with infectious keratitis. External validation included smartphone slit-lamp images from portable devices, simulating resource-limited environments in LMICs with infrastructural disparities.

Deep learning models, including convolutional neural networks (CNNs) and nnUNet frameworks, were trained on AS-OCT and slit-lamp images for tasks like anterior chamber angle classification, opacity quantification, and disease detection. Refraction correction and multitask learning segmented structures such as iris, lens, cornea, and trabecular meshwork. Models underwent five-fold cross-validation, with performance measured by mean intersection over union (MIoU), intraclass correlation coefficient (ICC), sensitivity, specificity, area under the curve (AUC), and mean absolute error (MAE).

AI models achieved MIoU of 0.959 for lens cortex and 0.928 for nucleus segmentation, ICC of 0.993 for opacity quantification, sensitivity/specificity of 0.999/1.000 for closed angle detection, and overall angle classification sensitivity/specificity of 0.989/0.995. For anterior chamber depth prediction from smartphone images, MAE was 0.093 ± 0.082 mm; diagnostic accuracy improved from 79.2% to 88.8% with AI assistance (p<0.001). CorneAI reached 86.0% classification accuracy across 100 images.