A multi-scale segmentation-to-classification network for tiny microaneurysm detection in fundus images

Microaneurysms (MAs) are recognized as the earliest indicators of Diabetic retinopathy (DR), which play an important role in DR screening. However, the detection of MAs is full of challenges due to the fact that the MAs are tiny spots in the complex situations. In this paper, we propose a multi-scale segmentation-to-classification model to improve the accuracy of d MAs detection in the complex situations. This proposed model is primarily composed of two stages: (1) the segmentation of MAs and; (2) the classification for MAs. In the segmentation stage, a multi-scale residual network, termed as MSRNet, is designed to explore the necessary information from the MAs with large size variance. In the classification stage, a multi-scale efficient network, named as MS-EfficientNet, is built to discriminate the false MAs from the MA candidates detected by the MSRNet. Furthermore, we reasonably expand the positive patches by adding constrained variances to the MA ground truths. Experiments show that the proposed model significantly outperforms the state-of-the-art with an average sensitivity score of 0.715 on the E-ophtha-MA datasets. The further results with other datasets also demonstrate that the proposed model for MAs segmentation is highly effective and robust. Our source code will be available on the GitHub: https://github.com/Lyblue11/Microaneurysm-Detection.