SUDS: A Simplified U-Net Architecture with Depth-Wise Separable Convolutions

Medical image segmentation is one of the most important topics in the field of computer vision and plays a crucial role in computer-aided diagnosis. U-Net paved the way for a series of variants that took advantage of the key characteristics of this network. In this article, several features proposed in different variants of U-Net are adapted and experimented upon to create a new architecture that maintains the idea of a U-shaped structure. The proposed architecture takes advantage of the efficient depth-wise separable convolution, but with a twist. Instead of using the pointwise convolution as the last step in the depth-wise separable convolution, it utilizes the so-called Ghost Module. This results in a highly efficient network with a reduced complexity, that still has excellent segmentation performance. We compared SUDS with U-Net and its variants across multiple segmentation tasks from two categories, skin lesion segmentation and colonscopy segmentation. Experiments demonstrate that SUDS has similar segmentation accuracy compared to the other networks, while the number of parameters and floating-point operations are greatly reduced.