Predicting response of anti-VEGF therapy in eyes with diabetic macular edema using deep-leaning based OCT image analysis: a multi-center study

Abstract

Purpose : Even anti-VEGF therapy is considered as the first line in managing center-involved diabetic macular edema (CI-DME), frequent injections are costly and even ineffective in some patients. We developed an integrated deep-learning (DL) system based on pretreatment OCT images and clinical variables to predict the therapeutic response in eyes with CI-DME after 3 monthly injections.

Methods : A total of 705 pre-treatment OCT volumes (16,477 B-scans) captured by Spectralis OCT (Heidelberg Engineering, Germany), representing 457 CI-DME eyes with 3 intravitreal injections of anti-VEGF agent from 366 patients with diabetes were retrospectively drawn from 3 retina clinics in Hong Kong for model training, tuning, and validation. Good response was defined as ≥ 1-line gain on the Snellen visual acuity (VA) chart or >10% reduction in central subfield thickness (CST). No response was defined as less than 1-line gain on the Snellen VA chart and ≤10% reduction in CST after 3 months of treatment. The DL system consisted of 3 networks (Figure): 1) a segmentation network based on a modified DeepLabv3+ architecture to segment 4 retinal features (i.e., intraretinal cysts, subretinal fluid, disorganization of retinal inner layers, and outer-retinal defects) from 5 OCT B-scans within foveal area; 2) a classification network based on a DenseNet121 network with 4 individual fully connected layer to detect the presence of the 4 features; and 3) a multi-modal ResNet-18 network that combined the stacked raw OCT images, the results from the above 2 networks, and clinical variables (pre-treatment VA and CST) to classify eyes into “good response” or “no response” if treated with anti-VEGF. External testing was performed using 3 independent datasets from Australia and India.

Results : The integrated DL system achieved an AUROC of 0.731 (95% CI 0.619-0.844), sensitivity of 70.0% (95% CI 55.6-87.4), specificity of 70.2% (95% CI 50.5-79.3) and accuracy of 69.5% (95% CI 62.4-77.3) in the internal validation. In the external testing, the DL system achieved AUROCs of 0.677, 0.652 and 0.669, and accuracies of 76.3%, 78.1% and 68.4%, respectively.

Conclusions : Our proposed DL system may be useful as a clinical decision support tool to classify eyes with CI-DME into “likely” or “unlikely” to have good prognosis when treated with anti-VEGF agents for assisting therapeutic management decision.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.