Multi-Scale Attention U-Net for Optic Disc and Optic Cup Segmentation in Retinal Fundus Images

This research illustrates the Multi-Scale Attention UNet (MSA-UNet) architecture to develop the retinal fundus image segmentation employing effective capturing complex features. The research innovation points mainly include the continuous integration of multi-scale image convolution layer and attention gate. The strategic integration of the decoder path and the developed superior performance compared to existing art of segmentation networks. The decoder path is split into two major paths. One represents the integration of attention gates and multi-kernel convolution. Another indicates the development of double $3 \times 3$ convolution for up sampling. The proposed research design not only enables the network to capture the complex features efficiently, but also leveraging attention mechanisms overwhelm irrelevant areas that improve the segmentation precision. This research mainly includes the ORIGA and REFUGE datasets to realize the superior performance of MSA-UNet. It is evident that employing the proposed Multi-Scale Attention UNet (MSA-UNet) architecture, the retinal fundus image segmentation significantly improved. The research has practical implication in the medical image analysis, particularly in the retinal fundus image segmentation with maintaining high precision. As a result, research may contribute to the diagnosis and treatment of retinal diseases, such as glaucoma and diabetic retinopathy.