Fully automated analysis of retinal optical coherence tomography (OCT) may have promising utility in the diagnosis and prognosis of geographic atrophy in both research and real-life settings, according to research results published in Lancet Digital Health. Researchers say their model “achieves performance at a similar level to manual specialist assessment.”

To develop and validate the fully-automated deep-learning method to perform quantitative OCT segmentation of geographic atrophy, researchers conducted a validation study of retinal scans at a single center in London. 

Segmentation models of OCT images were collected through the international, multicenter FILLY trial (ClinicalTrials.gov NCT02503332). An external dataset was created from a retrospective cohort of patients who were seen at a tertiary referral center with 32 clinic sites. 


Continue Reading

The model development dataset included 5049 B-scans from 984 OCT volumes from 399 eyes of 200 patients (mean age, 79±7.4 years; 53% women) between 2015 and 2016; the dataset was randomly split at the patient level. An external validation dataset — including 884 B-scans from 192 OCT volumes from 192 eyes of 110 patients) was assembled from 2016 to 2019. These patients had similar age and gender as the model development cohort (mean age, 78.3±11.1 years, 53% women). 

Investigators developed deep-learning models for each constituent OCT feature: RPE loss, overlying photoreceptor degeneration, and hypertransmission. Correlation coefficients between model prediction and consensus grading were utilized to evaluate the model accuracy, with similar feature performance observed for each model on the external validation set: median DSCs were 0.95 for RPE loss, 0.96 for overlying photoreceptor degeneration, and 0.97 for hypertransmission. Agreement between the model prediction and consensus grading for the internal and external validation datasets were higher than intergrader agreement. 

A subcohort analysis of other macular pathologies — including neovascular age-related macular degeneration, pigment epithelial detachment, and epiretinal membrane — was performed. High performance was noted for each feature, and was retained when investigators considered a subcohort inclusive of only 1 eye and its corresponding volume. 

Researchers explored 2 approaches to segment geographic atrophy. In the binary detection of geographic atrophy within the B-scans, both approaches resulted in high F1 scores (0.94; 95% CI, 0.92-0.96) and accuracy (0.91; 95% CI, 0.89-0.93). 

In terms of geographic atrophy segmentation tasks, predictions of both approaches were highly correlated with consensus grading of the external validation dataset. Investigators saw higher mean and median DSC values, as well as higher ICC values, in approach 1 compared with approach 2. However, both approaches outperformed the observed agreement between human graders (median DSC, 0.96, 0.95, and 0.80, respectively). 

The high performance in detecting overlapping regions — including RPE loss, photoreceptor degeneration, and hypertransmission — was retained in OCTs with other retinal pathologies. 

Study limitations include a lack of generalizability to other OCT devices and limited algorithmic output. 

“Predictive performance to a standard similar to clinical experts was reported,” according to researchers. “Notably, high performance was retained when evaluated on an external validation dataset…showing potential for real-life, point-of-care clinical utility.”

“This method could support the management of patients with geographic atrophy and has the potential to facilitate the development of standardized clinical trial endpoints for research into therapy development,” the study says. 

Disclosure: This clinical trial was supported by Apellis Pharmaceuticals. Please see the original reference for a full list of authors’ disclosures. 

Reference

Zhang G, Fu DJ, Liefers B, et al. Clinically relevant deep learning for detection and quantification of geographic atrophy from optical coherence tomography: A model development and external validation study. Lancet Digit Health. Published online September 8, 2021. doi:10.1016/S2589-7500(21)00134-5