Atrous-Ghost Module in Half-UNet for Segmenting Retinal Vessels in Fundus Images

In medical image analysis, retinal blood vessels segmentation is essential for the diagnosis and treatment of a variety of retinal diseases. This study introduces Atrous-Ghost Half-UNet (AGH-UNet), an advanced architecture of the Half-UNet architecture designed for enhanced retinal vessel segmentation. Our approach incorporates two primary innovations: Firstly, we integrate Atrous convolutions into the ghost module of the Half-UNet framework, replacing standard convolutions. This modification expands the receptive field without compromising spatial resolution, improving the capacity of the model in detecting large and small blood vessels with high accuracy. Secondly, we implement an Enhanced Mixing Loss (EML) function, adapted from the D-UNet model. The EML function addresses class imbalance issues by combining a weighted focal coefficient with a conventional loss term. By incorporating this loss function into our model, we aim to improve segmentation accuracy, particularly in scenarios involving imbalanced vessel structures or small vessels that are typically challenging to identify. Experimental results demonstrate that our proposed method achieves superior segmentation performance compared to the standard Half-UNet architecture. This research contributes to advancing retinal vessel segmentation techniques, potentially enhancing the efficacy of diagnostic and treatment processes for retinal pathologies.