Hybrid deep learning models for the screening of Diabetic Macular Edema in optical coherence tomography volumes

Abstract

Several studies published so far used highly selective image datasets from unclear sources to train computer vision models and that may lead to overestimated results, while those studies conducted in real-life remain scarce. To avoid image selection bias, we stacked convolutional and recurrent neural networks (CNN-RNN) to analyze complete optical coherence tomography (OCT) cubes in a row and predict diabetic macular edema (DME), in a real-world diabetic retinopathy screening program. A retrospective cohort study was carried out. Throughout 4-years, 5314 OCT cubes from 4408 subjects who attended to the diabetic retinopathy (DR) screening program were included. We arranged twenty-two (22) pre-trained CNNs in parallel with a bidirectional RNN layer stacked at the bottom, allowing the model to make a prediction for the whole OCT cube. The staff of retina experts built a ground truth of DME later used to train a set of these CNN-RNN models with different configurations. For each trained CNN-RNN model, we performed threshold tuning to find the optimal cut-off point for binary classification of DME. Finally, the best models were selected according to sensitivity, specificity, and area under the receiver operating characteristics curve (AUROC) with their 95% confidence intervals (95%CI). An ensemble of the best models was also explored. 5188 cubes were non-DME and 126 were DME. Three models achieved an AUROC of 0.94. Among these, sensitivity, and specificity (95%CI) ranged from 84.1-90.5 and 89.7-93.3, respectively, at threshold 1, from 89.7-92.1 and 80-83.1 at threshold 2, and from 80.2-81 and 93.8-97, at threshold 3. The ensemble model improved these results, and lower specificity was observed among subjects with sight-threatening DR. Analysis by age, gender, or grade of DME did not vary the performance of the models. CNN-RNN models showed high diagnostic accuracy for detecting DME in a real-world setting. This engine allowed us to detect extra-foveal DMEs commonly overlooked in other studies, and showed potential for application as the first filter of non-referable patients in an outpatient center within a population-based DR screening program, otherwise ended up in specialized care.

Keywords: Deep learning; Diabetic Macular Edema; Diabetic Retinopathy; Optical Coherence Tomography; Screening; Telemedicine.

Figure 1

AUROC and pAUROC (95%CI) for the classification of DME, with the best three models. ROC: receiver operating characteristics; AUROC: area under the ROC curve; DME: diabetic macular edema. Vertical dotted and colored lines represented pAUROC at FPR (1-specificity) < 0.05 and FPR < 0.10.

Figure 2

Characteristics of false positives and negatives obtained with the ensemble model at threshold 3. ERM: epiretinal membranes; DME: diabetic macular edema; AMD: age-related macular degeneration; PDR: proliferative diabetic retinopathy. ^*Includes ischemia and thinning, asteroid hyalosis, central serous choroidopathy, lamellar macular hole, vascular tortuosity, and epithelial pigment detachment.

Figure 3

ROC and AUROC (95%CI) for different grades of diabetic retinopathy with the best three models. ROC: receiver operating characteristics; AUROC: area under the ROC curve; DR: diabetic retinopathy; DME: diabetic macular edema. ^*Referable DR included severe non-proliferative DR, proliferative DR, moderate DME, and severe DM.