Machine learning accurately identifies iRORA and cRORA lesions in AMD

Srinivas R. Sadda, MD

Courtesy of UCLA

A new investigation aimed to evaluate a deep learning algorithm to automatically detect incomplete retinal pigment epithelium and outer retinal atrophy (iRORA) and complete retinal pigment epithelium and outer retinal atrophy (cRORA) in eyes with age-related macular degeneration (AMD).

The results suggest that a deep learning model can both accurately and precisely identify iRORA and cRORA lesions in optical coherence tomography (OCT) volumes of eyes with non-neovascular AMD.¹

“The model can achieve similar sensitivity to human graders, potentially avoiding a laborious and time-consuming annotation process and could be developed into a diagnostic screening tool,” wrote the survey team, led by Srinivas. R. Sadda, MD, Doheny Eye Institute, University of California, Los Angeles.

As part of the retrospective machine learning analysis, a deep learning model was trained to jointly classify the presence of iRORA and cRORA in a given B-scan. Then, the algorithm was evaluated using 2 separate and independent data sets. OCT B-scan volumes were captured from a total of 71 patients with non-neovascular AMD at the Doheny-UCLA centers.

Additionally, the following 2 external OCT B-scan test datasets were used: University of Pennsylvania, University of Miami and Case Western Reserve University and (2) Doheny Image Reading Research Laboratory.

The images were then annotated by an experienced grader for the presence of iRORA and cRORA. The investigators trained a Resnet18 model to classify the annotations for each B-scan using the OCT volumes collected from the Doheny-UCLA Eye Centers. This model was applied to 2 sets of test data to assess the performance of the out-of-sample model, according to the investigators.

The main results of the analysis were the performance of the model, quantified in terms of area under the receiver operating characteristic curve (AUROC) and area under the precision recall curve (AUPRC). The team further compared sensitivity, specificity, and positive predictive value against other clinician annotators.

After analysis, on an independently collected test set consisting of 1117 volumes from the general population, the investigators found that the model predicted the presence of iRORA and cRORA in the entire volume with AUROC performance and AUPRC scores almost perfect (iRORA, 0.61; 95% confidence interval [CI], 0.45 – 0.82; cRORA, 0.83; 95% CI, 0.68 – 0.95).

On another independently collected set, consisting of 60 enriched OCT B-scans for iRORA and cRORA lesions, the data showed the model performed with AUROC (iRORA, 0.68; 95% CI, 0.54 – 0.81 cRORA, 0.84; 95% CI, 0.75 – 0.94) and AUPRC (iRORA, 0.70; 95% CI, 0.55 – 0.86; cRORA, 0.82; 95%, 0.70 – 0.93).

Together, the researchers suggest that the results showed that a deep learning model is both accurate and precise in identifying iRORA and cRORA lesions in the OCT B-scan volume.

The references

1. Chiang JN, Corradetti G, Nittala MG, et al. Automated identification of incomplete and complete retinal pigment epithelial and outer retinal atrophy using machine learning. Ophthalmology Retina. 2023;7(2):118-126. doi:10.1016/j.oret.2022.08.016