Improved fuzzy C-means clustering in the process of exudates detection using mathematical morphology

• Published in: Soft computing (Berlin, Germany) Vol. 22; no. 8; pp. 2753 - 2764
• Main Authors: Wisaeng, Kittipol, Sa-ngiamvibool, Worawat
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.04.2018; Springer Nature B.V
• Subjects: Artificial Intelligence; Clustering; Computational Intelligence; Control; Developing countries; Diabetes; Diabetic retinopathy; Digital imaging; Engineering; Exudation; Human error; Image analysis; Image contrast; Image enhancement; Image quality; LDCs; Mathematical analysis; Mathematical Logic and Foundations; Mathematical morphology; Mechatronics; Medical personnel; Methodologies and Application; Morphology; Neural networks; Retinal images; Robotics; Support vector machines; Thailand; Diabetic retinopathy; Digital retinal image; Support vector machine; Exudate detection; Mathematical morphology; Naïve Bayesian; Fuzzy C-means clustering
• ISSN: 1432-7643; 1433-7479
• DOI: 10.1007/s00500-017-2532-8

Exudates are a common complication of diabetic retinopathy and the leading cause of blindness in the developing countries, especially in Thailand. The digital retinal images are usually interpreted visually by an expert ophthalmologist in order to diagnose exudates. However, detecting exudates in a large number of the digital retinal images is mostly manual and very expensive in expert ophthalmologist time and subject to human errors. In this research, we propose a novel retinal image analysis for detecting exudates through image preprocessing methods, i.e., histogram matching, local contrast enhancement, median filter, color space selection, and optic disc localization. Our in-depth retinal analysis indicates that the overall image quality is sensitive to the quality score. In the detection process, the exudates are detected by using naïve Bayesian classifier, support vector machine, and fuzzy C-means clustering method. Afterward, the exudates extracted from fuzzy C-means clustering are used as input to the mathematical morphology to obtain the final exudates detection quality score. Additionally, the optimal parameters of the mathematical morphology will increase the accuracy of the results from merely fuzzy C-means clustering method by 12.05%. The combination of these methods demonstrated an overall pixel-based accuracy of 97.45% including 97.12% sensitivity and 97.89% specificity.