Prediction Of Diabetic Retinopathy With Pyramidal Structured Convolution And Rank-Based Lesion Candidate Estimation

Abstract

This paper proposes a diabetic retinopathy (DR) grading approach that uses pyramidal structured convolution (Py-CNN) with a rank-based lesion candidate estimation process for extracting the deep lesion descriptors. Initially, the proposed approach segments the possible lesion candidates after preprocessing using the HSV transform-based adaptive histogram equalization approach. The top????segmented lesion candidates are ranked based on the edge and energy components present in the lesion candidates. The pyramidal structured convolution has different sections of the convolutional filter where each section processes the ranked lesion candidates. The lower section of the pyramidal structured convolution has higher length filters that process the top-ranked lesion candidates. The length of the convolutional filter reduces as the rank of the lesion candidate reduces. Two networks namely Network-Aand Network-B are utilized to extract the global fundus image and deep lesion candidature descriptors. The deep lesioncandidaturedescriptorsare implantedinthelesionregionofthe featuremap withthe use of a diffusion layer. The diffused descriptors are used by one of the dense networks, while the global descriptors and the deep lesion candidature descriptors are used by another dense network. Based on the predicted probability of the two dense networks the actual classification result can be computed. Datasets namelyAPTOS 2019 and Messidor-2 are utilized to evaluate

thealgorithmperformancewithdifferentevaluationscales.Theproposedapproachresultsin a specificity, Mathew’s correlation coefficient, F1-score, precision, and accuracy of 99.36%, 96.77%, 97.41%, 97.35%, and 97.47%when evaluated using the APTOS2019 datasets.