Automatic microaneurysms detection on retinal images using deep convolution neural network

Visual loss can be prevented by early detection and treatment of disease. Diabetic retinopathy is the leading cause of vision loss, and microaneurysms (MAs) are an early symptom of this disease. The fundus examination is effective at early detection of diabetic retinopathy. However, detecting MAs on retinal images is difficult for physicians because MAs typically appear as small dark dots. Therefore, many studies on automated MA detection have been conducted. This study itself proposes an MA detector that combines three existing types of detectors: the double-ring filter, shape index based on the Hessian matrix, and Gabor filter. However, because deep convolutional neural networks (DCNN) have shown superior performance in image recognition studies, this study conducts automated MA detection using DCNN. The proposed method is structured with a two-step DCNN and three-layer perceptron with 48 features for false positives (FPs) reduction. In the two-step DCNN, the first DCNN is for initial MA detection and the second DCNN is for FPs reduction. By applying the proposed method to the DIARETDB1 database, the proposed method shows superior performance.