Automatic microaneurysm detection in fundus image based on local cross-section transformation and multi-feature fusion
Jingyu Du, Beiji Zou, Changlong Chen, Xuzi Wen, Qing Liu
Abstract:
Background and objective: Retinal microaneurysm (MA) is one of the earliest clinical signs of diabetic retinopathy(DR). Its detection is essential for controlling DR and preventing vision loss. However, the spatial scale of MA is extremely small and the contrast to surrounding background is subtle, which make MA detection challenging. The purpose of this work is to automatically detect MAs from fundus images. Methods: Our MA detector involves two stages: MA candidate extraction and classification. In MA candidate extraction stage, local minimum region extraction and block filtering are used to exploit the regions where possibly exist MAs. In this way, most MAs are contained in candidates and many unimportant background regions are automatically filtered out, which facilitates the training of MA classifier in the second stage. In the second stage, multiple features are extracted to train the MA classifier. To distinguish MA from vascular regions, we propose a series of descriptors according to the cross-section profile of MA. Specially, As the MAs are small and their contrast to surroundings is subtle, we propose local cross-section transformation (LCT) to amplify the difference between the MA and confusing structures. Finally, an under-sampling boosting-based classifier (RUSBoost) is trained to determine a candidate whether an MA or not. Results: The proposed method is evaluated on three public available databases i.e. e-ophtha-MA, DiaretDB1 and ROC training set. It achieves high sensitivities for low false positive rates on the three databases. Using the FROC metric, the final scores are 0.516, 0.402 and 0.293 respectively, which is comparable to existing state-of-the-art methods. Conclusions: The proposed local cross-section transformation enhances the discrimination of descriptors by amplifying difference between MAs and confusing structures, which facilitates the classification and improves the detection performances. With the powerful descriptors, our method achieves state-of-the-art performance on three public datasets consistently.
Result:
Table 1 Sensitivities at predefined FPIs for different methods on the e-ophtha-MA
Table 2 Sensitivities at predefined FPIs for different methods on the DiaretDB1
Author :
Jingyu Du (jingyudu@csu.edu.cn), Beiji Zou (bjzou@csu.edu.cn), Changlong Chen (chen_c_long@sina.com), Xuzi Wen(csuwenwen@163.com), Qing Liu(qing.liu.411@gmail.com)