Cone-Rod Dystrophy Multimodal Prediction

Abstract

Cone-Rod Dystrophy (CRD) is a rare inherited retinal disorder that leads to the progressive degeneration of cone and rod photoreceptors, causing loss and worsening of visual functions, color defects, and visual field loss. Current diagnosis is based on a structural assessment by Optical Coherence Tomography (OCT) and the functional assessment by Electroretinography (ERG). CRD manifestations are subtle and variable in the initial stages, and unimodal diagnostics are not that sensitive. This paper proposes a multimodal diagnostic framework that combines deep learning and OCT data with machine learning and ERG data to automatically detect CRD. Eight deep learning models (ResNet18, ResNet-50, DenseNet-121, VGG16, InceptionV3, SE-ResNeXt50, Swin Transformer, and DeiT) were analyzed on OCT images; OCT segmentation on images was performed using YOLOv11. In addition to logistic regression, Support Vector Machines, Random Forests, and XGBoost were applied to identify the trends in 30dimensional ERG data (scotopic and photopic measurements). XGBoost and Swin Transformer were the best OCT and ERG classifiers, weighted at the decision level to produce a coherent output. The explainable AI is incorporated in the system using Grad-CAM images, which produce detailed diagnostic outputs that encompass OCT attention maps, ERG feature analysis, and clinical recommendations. Experimental validation of the study, which was done on 1,000 subjects (400 with CRD and 600 controls), gave an accuracy of 98.8 %, precision of 100 %, and a recall of 97.4 %. This proves that it is a clinically viable screening tool to detect CRD at its early phases.