A novel diabetic retinopathy grading using modified deep neural network with segmentation of blood vessels and retinal abnormalities

This paper proposes a new DR grading for solving the aforementioned problems. The initial process of the proposed model is the pre-processing, which is performed by median filtering. The segmentation of blood vessels and retinal abnormalities like exudates, microaneurysm, and hemorrhages is also done for grading the DR. Before initiating the segmentation of those components, optic disc removal is opted using the open-close watershed transform. Once the optic disc is removed, adaptive active contour methodology is used for performing the blood vessel segmentation. As a major contribution, the threshold value of the active contour method is optimized using proposed FNU-GOA with the aim of maximizing the accuracy of the blood vessel segmented images. Further, the retinal abnormalities like exudates, microaneurysm, and hemorrhages are performed by Otsu thresholding with morphological operation. From both segmented blood vessels and retinal abnormalities, the features like “Gray-Level Co-occurrence Matrix (GLCM)”, “area of Region of Interest (RoI)” and “Local Ternary Pattern (LTP)” are extracted. These features are subjected to the Modified Deep Neural Network (MDNN) for grading the DR. This MDNN focuses on solving the over fitting problems of DNN with the aim of maximizing the accuracy in terms of grading. The improvement of adaptive active contour-based blood vessel segmentation and MDNN-based grading is progressively dependent on the proposed Fitness-based Newly Updated Grasshopper Optimization Algorithm (FNU-GOA). This new algorithm improves the efficiency of grading with better convergence results. The experimental results show promising results as compared with other systems when analyzing various performance measures.