Detection of cystoid macular edema in patients with retinitis pigmentosa based on deep learning

Cystoid macular edema (CME) is a leading cause of vision loss in patients with retinitis pigmentosa (RP). The present study was aimed to CME in patients with RP using deep learning (DL) models based on the analysis of the optical coherence tomography (OCT) images.

In this cross-sectional study, a total of 1,318 OCT scans of 296 eyes from RP patients were analyzed with scans grouped based on the presence (670 images) or absence (648 images) of CME. We used Spectral-Domain OCT (SD-OCT) to measure central foveal thickness and detect retinal abnormalities, including subclinical CME in the study groups. The dataset was stratified and divided into training and testing sets using a subject-wise split at an 80:20 ratio using the scikit-learn library. Resnet-34 and ResNet-18 model architectures were developed to automatically detect CME in RP patients, and their performance was evaluated and compared with other DL algorithms.

Fine-tuning pretrained ResNet-34 and ResNet-18 models achieved an accuracy of 99.25%, 98.75%, F1-score of 99.26%, 98.77% and ROC of 99% and non-pretrained ResNet-34 and ResNet-18 achieved an accuracy of 80.64%, 82.13%, F1-score of 83.88%, 84.87% and ROC of 80%, 82% in detection of CME in RP patients.

This study was the first to apply DL algorithms to diagnose and manage CME in RP patients using OCT images. Pretrained ResNet models, particularly ResNet-34 with 99.25% accuracy, outperformed non-pretrained ResNet-34 with 80.64% accuracy. These results underscore the potential of pretrained models to aid in detection CME and supporting remote healthcare services.