Abstract
Purpose :
To develop and train a deep learning-based OCT classification algorithm for detecting disorganization of retinal inner layers (DRIL) in patients who have diabetes.
Methods :
HD OCT Raster scans from Cirrus HD-OCT (Carl Zeiss Meditec) were obtained from a total of 417 eyes from 229 patients at the Cleveland Clinic Cole Eye Institute. Raster scan used included horizontal (~90%) with several vertical (~10%) scans. We performed a retrospective study with OCT scans (2,392 B-scans) of 417 eyes with diabetes. The OCT scans were evaluated and graded for presence of DRIL by two graders with the third (supervisory) grader resolving disagreements. A total of 1354 B-scans used were labelled without DRIL, while 1038 were labelled with DRIL. The final grading of scans was used to label the training data for the convolution neural network (CNN) using transfer-learning protocol. The Matlab programing platform was used with the deep learning toolbox for training of modified DL based neural network (Alexnet).
Results :
On a single CPU system, the transfer learning training took <35 mins. The labelled data was divided 80/20 for training/testing purpose. With this training, our deep learning network was able to predict the presence of DRIL in new OCT scans with a high accuracy of 94.36 %. Receiver Operating Characteristic (ROC) curve for the classifier performance is shown in Fig. 1. The area under the ROC curve was 0.9880.
Conclusions :
Disorganization of retinal inner layers (DRIL) strongly correlates with visual acuity in patients with diabetic macular edema. Manual detection of DRIL is time consuming and often challenging due to the subjective nature of DRIL assessment. The present study demonstrates that deep learning based OCT classification algorithm can be used for a quick and accurate check for presence of DRIL. This developed tool can assist in clinical OCT grading, clinical research, and clinical decision-making.
This is a 2020 Imaging in the Eye Conference abstract.