Detection of Diabetic Retinopathy using Splat Feature Classification in Fundus Image

Abstract

Automated detection of retinal hemorrhages in fundus image [2] is crucial step towards early detection or screening is difficult among large population. A novel splat feature classification method is introduced to detect retinal hemorrhages. Classification is been achieved through supervised learning approaches. The performance of sensitivity and specificity is been improved while processing with retinal hemorrhages than with lesions. An area under receiver operating characteristics curve (ROC) of 0.96 can be achieved at splat level and 0.87 at image level. General Terms : Supervised Classification

CLRD: Collaborative Learning for Retinopathy Detection Using Fundus Images

“…The continuous development of artificial intelligence technologies, such as machine learning [ 4 , 5 , 6 , 7 , 8 ] and deep learning [ 9 , 10 , 11 ], has made the high-performance detection of retinopathy possible. Traditional machine learning methods have been widely applied in this field.…”

“…Traditional machine learning methods have been widely applied in this field. For example, Latha et al [ 4 ] introduced an efficient Splat feature classification method for detecting retinopathy features, including hemorrhages. This method improved data usability by performing operations like denoising, morphological processing, and dynamic thresholding on fundus image data.…”

“…Despite achieving superior classification performance, many retinopathy detection algorithms still fall short of meeting clinical practice requirements regarding diagnostic ACC and precision (PRE). Traditional image feature extraction algorithms can extract targeted lesion features but require professional medical knowledge [ 4 , 5 , 6 , 7 , 8 ]. Conversely, traditional direct image feature extraction algorithms have a stronger generalization ability but cannot accurately detect the specific scope of lesions, resulting in insufficient diagnostic ACC [ 4 , 5 , 6 , 7 , 8 ].…”

“…Traditional image feature extraction algorithms can extract targeted lesion features but require professional medical knowledge [ 4 , 5 , 6 , 7 , 8 ]. Conversely, traditional direct image feature extraction algorithms have a stronger generalization ability but cannot accurately detect the specific scope of lesions, resulting in insufficient diagnostic ACC [ 4 , 5 , 6 , 7 , 8 ]. Artificial intelligence diagnostic algorithms based on neural networks require a large number of labeled sample data and massive computing resources [ 9 , 10 , 11 ].…”