Author: Enrico Borrelli (Italy)
Co-authors: Jonathan D. Oakley, Domenico Grosso, Federico Borghesan, Costanza Barresi, Francesco Bandello, Giuseppe Querques
Exudative neovascular age-related macular degeneration (AMD) represents a leading cause of vision loss in older populations. This disease may be characterized by the exudation of fluid into the sub-retinal pigment epithelium (RPE), subretinal and intraretinal locations. Indeed, the presence and amount of intra- (IRF) and sub-retinal (SRF) fluids were demonstrated to be associated with visual prognosis in these patients. Neovascular pigment epithelium detachment (nPED) measurements have additionally been associated with visual outcomes in exudative neovascular AMD, but manual quantification using structural optical coherence tomography (OCT) is extremely time consuming with variable interpretation. Deep learning–based segmentation of volumetric OCT data may, however, represent a reliable and fast approach to automatically detect and quantify such anatomic features. The aim of this study was to validate a deep learning algorithm for automated IRF, SRF and nPED segmentations in eyes with exudative neovascular AMD and understand its performance relative to different graders.
Retrospective study at San Raffaele Scientific Institute, Milan, Italy.
This IRB-approved retrospective analysis used OCT data from 50 patients with exudative neovascular AMD collected using a Heidelberg Spectralis HRA+OCT device (Heidelberg Engineering, Heidelberg, Germany). Each OCT volume was separately labeled by two masked readers, R1 and R2, to delineate and quantify the amount of IRF, SRF and nPED. This cohort of 50 volumes was randomly divided into two subgroups: 35 for model training and 15 for evaluation. Two models, A1 and A2, were created based on gradings from readers R1 and R2, respectively. Area under the curve (AUC) values gauged detection performance, and quantification between readers and models was evaluated using Dice and correlation (R2) coefficients.
The deep learning–based algorithms had high accuracies for all fluid types between all models and readers: per B-scan IRF AUCs were 0.953, 0.932, 0.990, 0.942 for comparisons A1-R1, A1-R2 , A2-R1 and A2-R2, respectively; SRF AUCs were 0.984, 0.974, 0.987, 0.979; and nPED AUCs were 0.963, 0.969, 0.961 and 0.966. Similarly, and for the same comparisons, the R2 coefficients for IRF were 0.973, 0.974, 0.889 and 0.973; SRF were 0.928, 0.964, 0.965 and 0.998; and nPED were 0.908, 0.952, 0.839 and 0.905. The Dice coefficients for IRF averaged 0.702, 0.667, 0.649 and 0.631 again for comparisons A1-R1, A1-R2, A2-R1 and A2-R2, respectively; for SRF were 0.699, 0.651, 0.692 and 0.701; and for nPED were 0.636, 0.703, 0.719 and 0.775. In an inter-observer comparison between manual readers R1 and R2, the R2 coefficient was 0.968 for IRF, 0.960 for SRF, and 0.906 for nPED, with Dice coefficients of 0.692, 0.660 and 0.784 for the same features. Taken together, this uniformly demonstrates that the deep learning models were accurate in segmenting anatomic features in neovascular AMD eyes with performance akin to the human graders.
Our deep learning-based method applied on patients with exudative neovascular AMD can volumetrically segment structural metrics on OCT scans with human levels of performance. Our models also showed high accuracy when compared to gradings performed by a different reader to that used to build the model, further underlining the method’s validity and consistency. Once replicated in future, larger studies, our approach may prove to be a useful tool for evaluating eyes with neovascular exudative AMD.