DRSegNet: A cutting-edge approach to Diabetic Retinopathy segmentation and classification using parameter-aware Nature-Inspired optimization

Abstract

Diabetic retinopathy (DR) is a prominent reason of blindness globally, which is a diagnostically challenging disease owing to the intricate process of its development and the human eye's complexity, which consists of nearly forty connected components like the retina, iris, optic nerve, and so on. This study proposes a novel approach to the identification of DR employing methods such as synthetic data generation, K- Means Clustering-Based Binary Grey Wolf Optimizer (KCBGWO), and Fully Convolutional Encoder-Decoder Networks (FCEDN). This is achieved using Generative Adversarial Networks (GANs) to generate high-quality synthetic data and transfer learning for accurate feature extraction and classification, integrating these with Extreme Learning Machines (ELM). The substantial evaluation plan we have provided on the IDRiD dataset gives exceptional outcomes, where our proposed model gives 99.87% accuracy and 99.33% sensitivity, while its specificity is 99. 78%. This is why the outcomes of the presented study can be viewed as promising in terms of the further development of the proposed approach for DR diagnosis, as well as in creating a new reference point within the framework of medical image analysis and providing more effective and timely treatments.

Fig 1. General DR lesions observed in the fundus images.

Fig 2. Stages of development of diabetic retinopathy.

Fig 3. Proposed research methodology for DR diagnosis.

Fig 4

(a) Retinal Lesions in Fundus Images, (b) hard exudates, (c) soft exudates, (d) haemorrhages, and (e) microaneurysms.

Fig 5. GauGAN architecture.

Fig 6. KCBGWO flowchart.

Fig 7. FCEDN architecture.

Fig 8. Overview of feature extraction with fusion approaches and the FuNet model.

Fig 9. ELM configuration.

Fig 10. Detailed portrayal of the diabetic retinopathy identification workflow.

(a) Initial retinal image, (b) Enhanced image post-preprocessing, (c) Identification of red lesions (RL), (d) Identification of bright lesions (BL), (e) Quantitative feature labels on bright lesions, and (f) Quantitative feature labels on red lesions.

Fig 11. Feature extraction-based classification results using pre-trained CNN models.

Fig 12. AUC for each stage.