Bringing AI to Automatic Diagnosis of Diabetic Retinopathy from Optical Coherence Tomography Angiography

Abstract

Artificial Intelligence (AI) is significantly gaining interest in the field of Diagnostic and Functional Optical Imaging. As cutting-edge algorithms for decision-making are vast and medical imaging machines are diverse, the choice of the ultimate algorithm remains challenging. As a breakthrough in the field, our aim is to explore the adequate machine and deep learning algorithms that improve the classification of Optical Coherence Tomography Angiography (OCTA) Images, between normal and Diabetic Retinopathy (DR) images. The target was to provide an automatic paradigm for the medical staff to detect the presence of DR Lesions from OCTA images for diagnostic and monitoring purposes. Data were collected prospectively over a year from a comprehensive medical center in Lebanon. The mixed Convolution Neural Network (CNN)-Support Vector Machine Network (CNN, SVM) algorithm was utilized in the new paradigm and compared to the feed forward backpropagation NN, to the SVM and to the modified SVM. Results were evaluated independently for the presence or absence of DR using statistical metrics. Experimental results showcased promising association of deep learning to the early diagnosis of DR. Results manifested the high performance of the new paradigm, where the mixed algorithm applied to the functional OCTA surpassed the performance of the feed forward backpropagation NN. The sensitivity of the mixed (CNN, SVM) algorithm was 22.22% higher than that obtained by the feed forward backpropagation NN. Moreover, the specificity of classification of DR from OCTA images using mixed (CNN, SVM) algorithm was 24.44% higher than that obtained by the feed forward backpropagation NN. The precision was 25.47% higher in the new paradigm than that obtained by the feed forward backpropagation network, and the accuracy was 23.35% higher in the mixed (CNN, SVM) than that obtained by the feed forward backpropagation NN. This high performance plays a massive role in improving the diagnosis of DR, and thus Healthcare system and processing of information. As a future prospect, we aim to consider more algorithms and variables in the diagnosis of DR from OCTA images.