Feature-transfer network and local background suppression for microaneurysm detection

Microaneurysm (MA) is the earliest lesion of diabetic retinopathy (DR). Accurate detection of MA is helpful for the early diagnosis of DR. In this paper, an efficient approach is proposed to detect MA, based on feature-transfer network and local background suppression. In order to reduce noise, a feature-distance-based algorithm is proposed to suppress local background. The similarity matrix of feature distances is calculated to measure the difference between background noise and retinal objects. Moreover, a feature-transfer network is proposed to detect MAs with imbalanced data. For each training process, the optimized weights and bias are transferred to the next training, until the optimal network is generated. Experimental results demonstrate that the proposed approach can accurately detect subtle MAs surrounded by complex background. Furthermore, the sensitivity values on the public datasets are up to 98.3%, 100%, 99.3%, 100%, 96.5%, respectively. The proposed approach outperforms the state-of-the-arts, in terms of the competition performance measure score.