Quantitative Optical Coherence Tomography Angiography Features For Objective Classification And Staging Of Diabetic Retinopathy
March 2020
Alam M, Zhang Y, Lim JI, Chan RVP, Yang M, Yao X. Quantitative Optical Coherence Tomography Angiography Features for Objective Classification and Staging OF Diabetic Retinopathy. Retina. 2020 Feb;40(2):322-332. doi: 10.1097/IAE.0000000000002373.
The evaluation and management of diabetic eye disease is undergoing a paradigm shift, and it could not come soon enough. With diabetic retinopathy a leading cause of blindness worldwide, and a preventable one at that, retina specialists are racing to improve care. Recent large clinical trials like PANORAMA have shown that not only is the disease preventable, but also that treatment makes reversal of the disease possible.
These recent milestones spotlight the importance of proper staging and prevention of disease progression. In parallel with advancements in technology, such as optical coherence tomography angiography (OCTA), as well as computer-aided classification with machine learning, we are positioned to fulfill the promise of reducing blindness globally.
This study builds on prior quantitative OCTA work by the authors, and analyzes Angioview OCTA images using a support vector machine (SVM) classifier model to quantify and stage non-proliferative diabetic retinopathy (NPDR). 60 patients with mild, moderate, and severe NPDR and 20 control patients were included for a total of 120 NPDR and 40 control 6 x 6 mm OCTA images. Each OCTA image was evaluated for 6 features: blood vessel tortuosity, blood vascular caliber, vessel perimeter index, blood vessel density, foveal avascular zone area (FAZ-A), and foveal avascular zone contour irregularity (FAZ-CI).
The computer-aided classification was tested for its sensitivity, specificity, and accuracy in NPDR staging. Blood vessel density was the most accurate of the 6 features, having 93.89% for control versus all NPDR and 90.89% for control versus mild NPDR in a binary classification. Combining all OCTA features improved accuracy to 94.41% for control versus all NPDR and 92.96% for control versus mild NPDR in a binary classification. The SVM classifier model was 83.94% accurate in identifying control versus mild, moderate, and severe NPDR in a multiclass classification. Importantly, the temporal perifoveal region (FAZ-A and FAZ-CI) was found to be the most sensitive region for early detection of diabetic retinopathy.
Overall this study demonstrates how a computer-aided classification system may provide an objective and accurate way to screen and stage diabetic retinopathy. Analysis of the temporal perifoveal region may be particularly useful to achieve higher level resolution and accuracy in the management of patients. Limitations include the sample size as well as finding a way to minimize poor acquisition and/or OCTA image artifact that affects analysis for real world patients. The authors lay exciting groundwork, and more work remains in implementing OCTA computer aided classification algorithms in a practical and efficient manner on a large scale.