A novel retinal image segmentation using rSVM boosted convolutional neural network for exudates detection. (July 2021)

• Record Type: Journal Article
• Authors: Ghosh, Swarup Kr; Ghosh, Anupam
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: Retinal image -- Convolutional neural network -- RSVM -- Optic disc segmentation -- Exudates -- FROC; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102785
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We have proposed a ranking support vector machine (rSVM) with convolutional neural network in deep learning field for the detection of diabetic retinopathy. The spatial features of the retinal images have been extracted from RGB channel and mapped into a single binary features plane by the computing of pixel by pixel score using rSVM. A deep convolutional neural network has been designed for the retinal image segmentation followed by automatic anomaly detection using morphological operations. The rSVM computes a score function which is more suitable for multi-level classification to binary features classification in order to reduce the overall execution time in the segmentation task. To validate the proposed model we have employed publicly available database and the performance of them outperform state-of-the-art algorithms. Abstract: Retinal image analysis is an emerging research field in ophthalmological disease diagnosis since falsely detected optic disc, fovea, and blood vessels have become essential levels for automated diagnosis practices. In this article, we introduce a novel retinal image segmentation based on ranking support vector machine (rSVM) with convolutional neural network in deep learning field for the detection of diabetic retinopathy. Firstly, the spatial features of the retinal images have been extracted from RGB channel and mapped into a single binary features plane by the computing of pixel by pixel score using rSVM. Thereafter, we have designed a deep convolutional neural network for the retinal image segmentation followed by automatic anomaly detection using morphological operations. The rSVM computes a score function which is more suitable for multi-level classification to binary features classification in order to reduce the overall execution time in the segmentation task. The CNN has been designed with rSVM to define a consistent feature label in the network that reduces the number of channels in the CNN which lead to fast convergence. As a consequence, we have achieved good segmentation accuracy such as 96.4%, 97% and 98.2% for three different databases through post processing steps in comparison with other existing model.