Deep ensemble model for diabetic retinopathy classification

Abstract

Diagnosing Diabetic Retinopathy (DR) throughout mass screening of diabetic is significant to stop vision loss in a noteworthy ratio of the working populace. Earlier discovery and quantification of disease progression are important to prevent future vision loss. DR is diagnosed by retinal image analysis. To aid healthcare professionals in the analysis of DR, it is important to classify images into different levels of DR grading severity. Accurate segment of blood vessels is essential to facilitate the diagnosis of systemic vascular diseases. Consequently, in field of medicinal imaging, automated segmentation of blood vessels in retina from fundus images has developed in fame. Some automatic segmentation techniques exist, but low complexity and high accurateness are difficult to maintain in bright lesions or red lesions or abnormal retinal images with variations in luminance and contrast. For ensuring that blood vessels are not incorrect for red lesion caused by DR, a blood vessel segmentation model is proposed by a Texture-Based Modified K-means Clustering method (TBMKC) where texture features are considered for clustering. Local data is obtained by Local Binary Pattern (LBP), center symmetric LBP and statistical and texture features by Gabor filter accordingly. Cluster centroids are fine-tuned by particle swarm optimization (PSO) techniques. Scale Invariant Feature Transform (SIFT) method find out specific key points that cannot be speckled by any scaling or rotation process. It is followed by local adaptive threshold for segmentation of blood vessel patterns. A binary morphological operation removes unnecessary artifacts and regions. From the analysis, the selected TBMKC+PSO system achieves superior accurateness for varied data sets. newline