Deep learning enabled hemorrhage detection in retina with DPFE and splat segmentation in fundus images

Abstract

The range of diabetics, hypertension, occlusions in vascular are rapidly increasing in the modern era. Adversarial effects of these diseases are the organ damage which is increasing from child to old age people. One of the severe damages involves the causes of eye diseases like diabetic eye disease or Diabetic Retinopathy (DR). Late prediction of DR may lead to permanent loss of vision. Hence there is a need among ophthalmologists in detecting and treating the eye diseases at the earliest stage. Detection of abnormalities as early as possible is a crucial task in today world as the existing strategies possess some setbacks. In this research work, a deep learning framework has been developed for the betterment in the prediction of retinal hemorrhage with the fundus image. The Double Pierced Feature Extraction (DPFE) is planned by merging the Enhanced Long Short-Term Memory (ELSTM) CNN and Maximally Stable Extremal Regions (MSER) algorithm. Splat algorithm is focused for segmentation which partitions a retinal image into various segments called splats of similar color, similar intensity and the spatial position. Every splat carries several details from that diverse features can be extracted. This splat based segment establishes a boundary based on an appropriate set of detailed information—needed features such as different area, filter, texture, splat related features and color etc. At last trained ELSTM with CNN is applied to predict the redder lesions of the fundus image, which confirms the presence of hemorrhage in the retina. The developed framework was evaluated using the DIARETDB2 dataset, and its performance was measured using parameters such as Classification Accuracy, Negative Acceptance Rate (NAR), Specificity, and Sensitivity and the experimental results highlight that the proposed methodology achieved an impressive sensitivity of 98.67% and specificity of 98.91%, outperforming existing techniques such as “ANFIS Classification with Cuckoo,” “ANFIS Classification with PSO,” and “Motion Pattern Recognition.” This analysis confirms the effectiveness of the proposed framework in accurately identifying relevant instances while excluding non-relevant patterns, showcasing its potential for improved pattern recognition and classification.