Automatic detection and segmentation of optic disc using a modified convolution network. (July 2022)

• Record Type: Journal Article
• Authors: Maiti, Souvik; Maji, Debasis; Kumar Dhara, Ashis; Sarkar, Gautam
• Is Part Of: Biomedical signal processing and control. Volume 76(2022)
• Publication Date: 2022-07
• Subjects: Optic disc -- Diabetic retinopathy -- Convolutional neural network; 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.2022.103633
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Diabetic Retinopathy refers to an eye disease commonly found in patients suffering from prolonged diabetes mellitus and can eventually cause blindness resulting from leaking of retinal blood vessels. Early detection of any abnormalities in the optic disc can help the ophthalmologists in effective treatment of the patient. Segmentation of optic disc is also necessary for screening glaucoma using optic cup-to-disc ratio. Automatic screening of fundus images using computerized techniques can be of great help in detection of several retinal pathologies. The study presents a deep learning approach for detecting and segmenting the optic disc automatically from the fundus images using a modified convolutional network. The most powerful encoder structure has been identified by experimenting on seven encoder networks namely VGG11, VGG13, VGG16, VGG19, Resnet34, Densenet121 and InceptionV3. In this paper, the VGG16 framework is considered as an encoder, while the decoder is designed as a symmetric structure of that of the encoder to perform object segmentation more effectively. The Convolutional Long Short Term Memory is incorporated in the encoder framework for better convergence and to obtain high accuracy. The efficiency of this work is verified by applying it over a large variation of fundus images from databases namely MESSIDOR, DRIVE, DIARETDB0, DIARETDB1, CHASE-DB1, IDRiD and STARE. Intersection-over-Union, Dice Coefficient, Accuracy and Sensitivity have been computed for evaluating the effectiveness of the work. This algorithm has attained an overall accuracy of 99.44%. The experimental results indicate the superiority of this model on segmentation operation.