DEC-DRR: deep ensemble of classification model for diabetic retinopathy recognition.

Abstract

Most diabetes patients are liable to have diabetic retinopathy (DR); however, the majority of them might not be even aware of the ailment. Therefore, early detection and treatment of DR are necessary to prevent vision loss. But, avoiding DR is not a simple process. An ophthalmologist can typically identify DR through an optical evaluation of the fundus and through the evaluation of color pictures. However, due to the increased count of DR patients, this could not be possible as it consumes more time. To rectify this problem, a novel deep ensemble-based DR classification technique is developed in this work. Initially, a Wiener filter (WF) is applied for preprocessing the image. Then, the enhanced U-Net-based segmentation process is done. Subsequent to the segmentation process, features are extracted that include statistical features, inferior superior nasal temporal (ISNT), cup to disc ratio (CDR), and improved LGBP as well. Further, deep ensemble classifiers (DEC) like CNN, Bi-GRU, and DMN are used to recognize the disease. The outcomes from DMN, CNN, and Bi-GRU are then subjected to improved SLF. Additionally, the weights of DMN, CNN, and Bi-GRU are adjusted via pelican updated Tasmanian devil optimization (PU-TDO). Finally, outputs on DR (microaneurysms, hemorrhages, hard exudates, and soft exudates) are obtained. The performance of DEC + PU-TDO for diabetic retinopathy is computed over extant models with regard to different measures for four datasets. The results on accuracy using the DEC + PU-TDO scheme for the IDRID dataset is maximum around 0.975 at 90th LP while other models have less accuracy. The FPR of DEC + PU-TDO is less around 0.039 at the 90th LP for the SUSTech-SYSU dataset, while other extant models have maximum FPR.