Abstract
Purpose :
To develop a reliable deep-learning algorithm for automated detection and quantification of different kinds of exudative manifestations associated with neovascular age-related macular degeneration (nAMD) in macular spectral-domain optical coherence tomography (SD-OCT) volume scans.
Methods :
107 SD-OCT cubes with active nAMD were extracted, out of which 92 were used for developing the automatic algorithm and the remaining 15 served as a testing set to evaluate the performance. In each B-scan the regions containing intra- and subretinal fluid (IRF, SRF), as well as retinal pigment epithelium detachment (PED) were manually annotated by the retina expert. The performance of the algorithm was measured in terms of detection and exact quantification of the above-mentioned fluids.
Results :
The fully-automated algorithm achieved good performance for quantification of IRF, SRF and PED, with the Pearson correlation coefficients between the estimated and manually annotated fluid volumes of 0.998, 0.994 and 0.984, respectively. In terms of pixel-wise segmentation metrics the algorithm achieved: sensitivity of 0.825%, 0.685% and 0.882% respectively, precision of 0.667%, 0.720%, and 0.794% and dice score of 0.721%, 0.674% and 0.834%. The example predictions are shown in Figure 1. The estimated fluid volumes were subsequently used to determine the presence/absence of pathology in the test B-scans, obtaining area under the curve (AUC) of 0.972, 0.966 and 0.986 for IRF, SRF and PED respectively.
Conclusions :
The fully automated deep learning model based on convolutional neural network showed good performance not only in terms of fluid detection but also quantification of pathological changes in nAMD, critical for longitudinal evaluation and monitoring. It therefore has the potential to provide a more consistent and detailed statistic describing nAMD patients and consequently lead to better reproducibility and more informed treatment decisions.
This is a 2020 ARVO Annual Meeting abstract.