Abstract
Purpose :
To present a deep learning-based method, trained on healthy eyes, to screen for eyes with retinal pathology using optical coherence tomography (OCT) scan data.
Methods :
The proposed deep learning-based anomaly detection and localization method uses deep convolutional generative adversarial network (DCGAN) and mean structural similarity (MSSIM) to reconstruct missing or occluded retinal layer and detect retinal anomalies from OCT scans. The proposed method has three steps. The first is to reconstruct missing or occluded regions of the retinal layer. The second step is to localize the anomaly and the third is to measure the degree or severity of the anomaly. The training data fed into the framework are all images from macular volumetric OCT scans of healthy subjects (DRI OCT Triton, Topcon Corp). The outputs from the framework are the volume label of normal versus abnormal and a volume score associated with the abnormality. To evaluate the accuracy, 79 OCT volumes from 28 normal eyes and 83 OCT volumes from 29 eyes having various retinal pathologies were fed into the framework as the test data sets.
Results :
The current anomaly detection method demonstrated high accuracy on both normal (97.5%) and disease groups (97.6%). The ROC area under the curve is 0.98. It is observed that the volume scores correlate with the degree of abnormal structural changes. Fig. 1 (a) and (b) have MSSIM scores of 0.607 and 0.899 respectively. The score is a measure of similarity in the range 0 to 1, higher tending to be normal and lower tending to be anomalous.
Conclusions :
The proposed anomaly detection method was able to produce reliable and accurate results and is promising with respect to practical implementation of an automated screening methodology for retinal abnormalities.
This is a 2020 ARVO Annual Meeting abstract.