Automated diabetic retinopathy grading and lesion detection based on the modified R-FCN object-detection algorithm

In this work, we develop a computer-aided retinal image screening system that can perform automated diabetic retinopathy (DR) grading and DR lesion detection in retinal fundus images. We propose a modified object-detection method for this task via a region-based fully convolutional network (R-FCN). A feature pyramid network and a modified region proposal network are applied to enhance the detection of small objects. The DR-grading model based on the modified R-FCN is evaluated on the Messidor data set and images provided by the Shanghai Eye Hospital. High sensitivity of 99.39% and specificity of 99.93% are obtained on the hospital data. Moreover, high sensitivity of 92.59% and specificity of 96.20% are obtained on the Messidor data set. The modified R-FCN lesion-detection model is validated on the hospital data set and achieves a 92.15% mean average precision. The proposed R-FCN can efficiently accomplish DR grading and lesion detection with high accuracy.