A novel preprocessing unit for effective deep learning based classification and grading of diabetic retinopathy

Abstract

Early detection of diabetic retinopathy (DR) is crucial as it allows for timely intervention, preventing vision loss and enabling effective management of diabetic complications. This research performs detection of DR and DME at an early stage through the proposed framework which includes three stages: preprocessing, segmentation, feature extraction, and classification. In the preprocessing stage, noise filtering is performed by fuzzy filtering, artefact removal is performed by non-linear diffusion filtering, and the contrast improvement is performed by a novel filter called Adaptive Variable Distance Speckle (AVDS) filter. The AVDS filter employs four distance calculation methods such as Euclidean, Bhattacharya, Manhattan, and Hamming. The filter adaptively chooses a distance method which produces the highest contrast value amongst all 3 methods. From the analysis, hamming distance method was found to achieve better results for contrast and Euclidean distance showing less error value with high PSNR. The segmentation stage is performed using Improved Mask-Regional Convolutional Neural Networks (Mask RCNN). In the final stage, feature extraction and classification using novel Self-Spatial Attention infused VGG-16 (SSA-VGG 16), which effectively captures both global contextual relationships and critical spatial regions within retinal images, thereby improving the accuracy and robustness of DR and DME detection and grading. The effectiveness of the proposed method is assessed using two distinct datasets: IDRiD and MESSIDOR.