CONTINUAL LEARNING-DRIVEN LAPLACIAN PALE TRANSFORMATIVE CONVOLUTION NETWORK FRAMEWORK FOR DIABETIC RETINOPATHY DETECTION FROM RETINAL FUNDUS IMAGES

Diabetic retinopathy (DR) is a diabetes-related eye ailment caused by retinal blood vessel (BV) damage. This
manuscript presents a novel continual DL framework for efficient DR disease detection. Initially, input retinal
fundus images are taken from the IDRI Dataset for accurate DR disease detection, which undergoes the
preprocessing stage by employing a Color Wiener Filter (CWF) that can enhance image clarity by adaptively
removing noise while maintaining edge details for further processing. After preprocessing, a novel Laplacian
Pale Transformative Convolution network (LPTCN) is introduced, which classifies with more accuracy the
distinction between different DR abnormalities. Moreover, the proposed framework integrates Elastic Weight
Consolidation (EWC) and Herding Selection Replay (HSR) to prevent catastrophic forgetting on new data
samples. The proposed framework is simulated in the Python platform. In the simulation part, the average
Accuracy of 97%, Matthew’s Correlation Coefficient (MCC) of 0.936, symmetric mean absolute percentage
error (SMAPE) of 2.07, and Computation Time (CT) of 4.9s, Youden’s index (YI) of 0.89 are obtained by
the suggested framework on DR identification.