MS-UNet: a lightweight Multi-Scale UNet for skin lesion segmentation

In recent years, transformer and convolutional neural networks (CNNs) have been widely employed in medical image segmentation due to their excellent feature extraction capabilities. However, the computational requirements imposed by the large number of model parameters have posed significant limitations on the practical applications of these models.In this paper, we propose a lightweight Multi-Scale UNet(MS-UNet), which significantly reduces the number of model parameters. Simultaneously, the model maintains high accuracy through the utilization of depth-wise separable convolutions and multi-scale fusion modules.Moreover, an improved pyramid convolution(PConv) is employed in encoders, while utilizing different strip convolutions in decoders to extract features of different scales.Experimental results demonstrate that MS-UNet achieves the performance of state-of-the-art (SOTA) methods with very low parameters and computation cost.Specifically, our MS-UNet has only 23.7K parameters and 78.3M GFLOPs on the ISIC2017 and ISIC2018 datasets, compared with EGE-UNet which is the first model with a parameter size within 50KB.