Microaneurysm detection in fundus images using small image patches and machine learning methods

Abstract

Diabetic retinopathy (DR) is an eye abnormality caused by long term diabetes and it is the most common cause of blindness before the age of 50. Microaneurysms (MAs), resulting from leakage from retinal blood vessels, are early indicators of DR, yielding a large body of diagnostic work focused on automatic detection of MA. However, automated detection of MAs is difficult because (1) the small size of MA lesions and low contrast between the lesion and its retinal background, (2) the large variations in color, brightness and contrast of fundus images, and (3) the high prevalence of false positives in regions with similar intensity values such as blood vessels, noises and non-homogenous background. In this paper, we analyzed MA detectability using small 25 by 25 pixel patches extracted from fundus images in the DIAbetic RETinopathy DataBase — Calibration Level 1 (DIARETDB1). Raw pixel intensities of extracted patches served directly as inputs into the following classifiers: a random forest (RF), a neural network (NN), and a support vector machine (SVM). We also explored the use of two techniques (principal component analysis and random forest feature importance) for input dimensionality. With traditional machine learning methods and leave-10-patients-out cross-validation, our method outperformed a deep learning based MA detection method [1], with AUC performance improved from 0.962 to 0.985 and F-measure improved from 0.913 to 0.926, using the same DIARETDB1 database.