Deep-learning-aided Detection of Referable and Vision Threatening Diabetic Retinopathy based on Structural and Angiographic Optical Coherence Tomography

Abstract

Purpose : We propose two fully automated deep-learning-aided frameworks for detecting referable and vision threatening DR (rDR and vtDR) from volumetric and en face structural and angiographic optical coherence tomography (OCT) data.

Methods : 3×3-mm macular OCTA scans were acquired from the eyes of 50 healthy participants and 293 patients with diabetes (with or without DR) using a spectral-domain OCTA system (Avanti RTVue-XR, Optovue Inc). Masked trained retina specialists graded the disease severity based on the Early Treatment of Diabetic Retinopathy Study (ETDRS) scale using 7-field fundus photography. rDR was defined as level 35 or worse, or any level with diabetic macular edema (DME). vtDR was defined as level 53 or worse, or level with DME. A 3D (EfficientNet-3D-B0) and a 2D (DcardNet-36) frameworks were constructed. Each was trained separately for rDR and vtDR. The rDR/vtDR detection performance of 3D and 2D frameworks were respectively optimized using the area under the receiver operating characteristic curve (AUC) and specificity at sensitivity over 95%. Class activation maps (CAMs) were also generated from each framework. Performance was evaluated with 5-fold cross-validation, with 60%, 20%, and 20% of the data split for training, validation, and testing, respectively.

Results : The 3D framework achieved AUC of 0.89±0.04 for rDR and 0.88±0.02 for vtDR. The 2D framework superior AUC of 0.95±0.02 for rDR (p=0.0130) and 0.94±0.02 for vtDR (p=0.0009). 2D framework demonstrated higher overall diagnostic accuracy, sensitivity, and AUC (t-test on 5-fold cross-validation output) (Table 1). By overlaying CAM on OCTA three-dimensionally, we found the regions with higher values in the CAMs were mostly focused on the non-perfusion areas (NPA) and fluid areas (Fig. 1).

Conclusions : We demonstrated that a deep-learning based framework can detect rDR and vtDR based on OCTA volume and 2D framework shows superior diagnostic power than 3D framework as trained by the currently available dataset.

This is a 2021 ARVO Annual Meeting abstract.