Classifying retinal degeneration using OCT and histological images in a rodent model for retinal degeneration by deep learning.

Abstract

Artificial intelligence (AI) is increasingly being applied in vision research for disease classification and outcome prediction. We present the first study to classify stages of retinal degeneration in a well-established preclinical model using both optical coherence tomography (OCT) and histological images, and to predict visual acuity from OCT images. The Royal College of Surgeon's (RCS) rat, a widely used and well-characterized model of retinal degeneration, was used. Starting postnatal day (P) 21, 35 rats of both sexes were tested for visual acuity, OCT, and histology at defined time points. OCT (n = 62,070) and histological (n = 16,306) images were split into training, validation, and testing sets. A ResNet18 model was trained and compared against human observer classification. Paired OCT and visual acuity data were used to train a predictive model. This study reveals that ResNet18 classify retinal degeneration stages from OCT images with 95.95 % accuracy (F1 = 94.93 %) and histological images with 90.71 % accuracy (F1 = 91.57 %), performing comparably to human observers. Visual acuity was predicted from OCT images with a mean squared error of 0.011 and mean absolute error of 0.069. Deep learning enables accurate and objective classification of retinal degeneration stages from multimodal imaging and can predict functional outcomes. This framework, validated in a preclinical model, establishes a path for AI-based, cross-species monitoring of retinal degeneration and therapeutic response over time.