Microaneurysm Detection on Retinal Fundus Images with Multi-Resolution GlobalNet to Support the Diagnosis of Diabetic Retinopathy

Diabetic retinopathy, one of the main causes of blindness in the world, can be diagnosed early by utilizing microaneurysms (MA). The existence of microaneurysms can be detected by performing semantic segmentation on the retinal fundus image. In this paper, deep learning method testing was carried out using the GlobalNet model at various input image resolution values to see the effect of resolution alteration on MA detection in retinal fundus images. The experiment provided AUPR values of $0.391 \pm 0.026,0.387 \pm 0.035$, and $0_{-}394\pm 0.050$, along with Fl scores of $0.361 \pm 0.034,0.361 \pm 0.022$, and $0.360 \pm$ 0.022. Evaluation done at pixel and lesion levels shows that the difference in resolution of the input images does not cause a significant change in the AUPR and Fl scores. However, the resolution alteration does affect the total number of false positive lesions in outputs. In addition, a combination method is developed, providing a better trade-off in sensitivity and precision compared to the GlobalNet segmentation model.