Abstract
Purpose :
To develop a common deep learning platform for diagnosis and classification of diabetic retinopathy (DR) using simultaneous fundus and OCT imaging of the retina
Methods :
The platform was trained with 4586 frames from OCT scans and 3165 fundus images of 629 DR patients screened at Narayana Nethralaya, Bangalore, India. Each frame was annotated by two retina specialists independently. A deep neural network screening model to classify each frame into normal and abnormal frames while simultaneously localizing corresponding pathologies was developed. Abnormal frames were indicated by the presence of any kind of traction, fluid filled regions (FFR), hard exudates or alteration in intraretinal layers. The localized regions were extracted as image patches and further classified into one of the above pathologies by a 2nd deep learning model (Pathology model). Further classification into non-proliferative DR (NPDR), proliferative DR (PDR) and non-referable was performed.
Results :
The platform for OCT achieved an area under the curve (AUC), sensitivity, specificity of 0.98, 95% and 95%, respectively. The “Pathology model” differentiated the classes with a precision, recall and F1 score of 88%, 88% and 0.88, respectively. The platform for fundus achieved an AUC, sensitivity and specificity of 0.97, 91% and 92%, respectively. Without prior consultation with the clinician, the platform achieved sensitivity and specificity of 93% and 72%, respectively.
Conclusions :
Deep learning network training was successful in detecting and identifying abnormal images of Asian-Indian patients. The platform achieved similar performance with both fundus and OCT images. Further, the similar performance between the two imaging modalities confirmed that either a fundus or OCT imaging may be adequate for screening of diabetic patients in primary care centers.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.