An Enhanced Hybrid Ensemble Deep Learning Framework for Automated Retinal Lesion Segmentation in Diabetic Retinopathy

Abstract

Diabetic retinopathy (DR) is one of the leading causes of visual disability and early diagnosis and monitoring of the disease is critical for accurate segmentation of retinal lesions. Automated segmentation, however, is difficult because of the variations of lesions characteristics and image quality of the retina. In this study, an Enhanced Hybrid Ensemble Deep Learning Model (HEDLM) is proposed in order to automatically segment the retinal lesions using the hybrid model of U-Net, U-Net+, and U-Net++. A dataset of DR2 retinal fundus images with 757 images was tested using the framework. The image preprocessing steps involved resizing, normalization, and augmentation of the data. The three segmentation models were trained separately with a composite Dice and Binary Cross-Entropy loss function and their results fused by averaging over probabilities, weighted fusion and voting on majority method. The proposed ensemble outperformed the other models such as U-Net, U-Net+ and U-Net++ and recorded Dice coefficient 0.90, Intersection over Union 0.84, and a Segmentation accuracy of 96%. The results show the effectiveness of complementary segmentation architectures for enhancing lesion localization and robustness of segmentation. The framework proposed is a computer-aided method for analysis of retinal lesions that could help facilitate screening for diabetic retinopathy (DR) and aid in clinical decision making.