SwinUnet with Multi-task Learning for Image Segmentation

Image segmentation finds extensive application in various scenarios. While transformer-based variant models have significantly enhanced image segmentation performance, the stability of model training remains a concern. To address these challenges, this study introduces the Multi-Task SwinUnet (MSwinUnet) framework. It achieves multi-task learning by incorporating an additional task, Mask Reconstruction Segmentation (MaskRSeg), alongside the original Image Segmentation (ImgSeg) task. Our approach seamlessly integrates with Swin-Unet, enhancing the model's segmentation performance. Extensive experimental results demonstrate that MSwinUnet surpasses baseline models including UNet, TransUNet, and Swin-Unet, achieving DSC of 89.53% and MIoU of 0.8176 on the ACDC benchmark dataset. Furthermore, optimal model stability is achieved when the task ratio for ImgSeg and MaskRSeg is 8:2. Furthermore, we thoroughly illustrate the variations in the effects of different mask rate and mask patch size parameters on the MaskRSeg task. Among these parameters, a mask rate of 45% and a mask patch size of 4 yield the most optimal segmentation results. The training approach proposed in this paper will assist in further improving the accuracy of image segmentation for a wider range of Transformer variant models.