GAPS-U-NET: Gating Attention And Pixel Shuffling U-Net For Optic Disc Segmentation In Retinal Images

Segmentation of the Optic Disc (OD) is a crucial procedure in detecting potential damage to the optic nerve in the retina. This is vital for the screening of Glaucoma and Diabetic Retinopathy (DR) which are the leading cause of irreversible loss of vision worldwide. However, manual segmentation is inefficient, expensive, time consuming and unsustainable. This paper presents GAPS-U-NET, a novel framework with the potential for revolutionizing clinical practice, especially in the context of early detection of Glaucoma and DR. It is a modified U-Net architecture that adopts CNN models as the backbone of the encoder path and applies the upsampling operation of periodic shuffling convolution with gating attention for fast convergence of the network and accurate segmentation. The proposed method overcomes three major challenges of existing encoder-decoder techniques for image segmentation, which are shallow networks, reduced feature resolution, and localization accuracy. First, we employ pretrained CNN model to gain the major benefits of very deep networks that can learn very complex functions to extract more discriminative features for OD segmentation. Then the gating attention block is added to the skip connections to ignore irrelevant regions and focus on important features of a specific region. Finally, pixel shuffling is added to improve feature resolution and segmentation accuracy. GAPS-U-NET is trained on the Indian Diabetic Retinopathy Image Dataset (IDRID) and yielded IOU of 92.83% and an accuracy of 99.71 % which shows that the proposed method can be incorporated into Glaucoma and DR screening programs. GAPS-U-NET is modular therefore it can be integrated into other medical image segmentation tasks.