A R-CNN Based Approach for Microaneurysm Detection in Retinal Fundus Images

摘要

Diabetic retinopathy (DR) is one of the major diseases causing blindness, and microaneurysms in the fundus are the first reliable lesions in its early stage. 'This paper proposes an object detection method for microaneurysms based on R-CNN, which consists of five steps: image preprocessing, candidate region generation, feature extraction, classification and non-maximal suppression. First, a fundus image preprocessing method and a candidate region generation algorithm for microaneurysms are proposed. Then, the VGG16 network is trained using the transferred fine-tuning model to extract features from candidate samples. Finally, real aneurysms are selected from candidate regions by a classifier. The experimental results in the internationally published database e-ophtha show that the proposed method outperforms other known related methods on the FROC indicator.