Pixel-level Diabetic Retinopathy Lesion Detection Using Multi-scale Convolutional Neural Network

Diabetic retinopathy (DR) is one of the leading causes of preventable blindness. It's urgent to develop reliable methods for auto DR screening, the key of which is the detection of lesions. This paper presents an innovative method to detect DR lesions in pixel-level. We design a multi-scale Convolution Neural Network (CNN) that make the full use of multiple different scales with complementary image information. Experiments are carried out on both private and public datasets. Results show that multi-scale CNN model outperforms single-scale CNN model and other state-of-the-art approaches.