Purpose
Microaneurysms (MAs) are said to be the first sign of diabetic retinopathy (DR). Counting the number of MAs is important for monitoring the progression of DR. Factors making MA detection a challenge include the variation in MA size, low and varying image contrast, uneven illumination and variation in fundus image background. Our objective is to design a fully automatic MA detection system with a high sensitivity and specificity.
Methods
MAs and blood vessels were selected by using local thresholding technique in green channel. Artefacts due to image background were identified by contrast, variance and standard deviation information; blood vessels were segmented by wavelet transform technique. Removal of background and blood vessels from the fundus images left initial MA candidates. Based on shape parameters (eccentricity, aspect ratio and circulatory) and calculated features (energy, standard deviation and variance), the candidates were filtered using Curvelet coefficients. To further reduce false positives, color features like intensity in HSV channels and standard deviation and variance in HSV and RGB channels were used.
Results
Validation was performed using publically available ‘Retinopathy Online Challenge (ROC)’ dataset. The dataset consists of 50 training images with ground truth and 50 test images. Out of those 50 training images, 37 images contain MAs (336 MAs in total). The proposed algorithm was evaluated with all those 37 images. Our method detected 125 MAs thus achieving 37.2 % sensitivity with 74.89 false positives per image (FPPI). The experiment result was compared with published results (refer to Table 1). Adal achieved better sensitivity and specificity by using a total of 64-SURF descriptors, three features from Radon space and two image patch features, in contrast to our 19 features (6 color features, 5 shape parameters and 8 statistical features). Lazar constructed peak map and extracted MA candidates using hysteresis thresholding. Regions of interest (ROIs) were manually selected. Our approach is however fully-automated.
Conclusions
A MA detection system based on local thresholding and Curvelet transform has been proposed. The system performance was evaluated with other existing techniques. Compared with the state-of-the-arts, it uses significantly less feature size (hence faster computation) and is fully automated while achieving comparable results.