| |
| Transformer, the prevalent backbone architecture in natural language processing, has been adopted in various vision tasks since the proposition of vision transformer. The performance of transformer has been proved to be almost the same as CNN's and even be better with large enough dataset. However, the initial vision transformer suffered from the straightforward structure, which requires large parameters and expensive computation cost, especially in the dense prediction task. This paper concentrates on medical image semantic segmentation task. In medical scene, UNet is always the popular backbone and many researchers proposed transformer-CNN or pure transformer UNet model recently. But the inherent feature misalignment caused by resizing feature maps and concatenation is still lack of focus. In this paper, a lightweight transformer-CNN hybrid UNet, SOFT-AlignUNet (SOFT-AU) , is proposed to solve above issues. On one hand, a novel softmax-free transformer, which reduces the calculation cost to be linear to the patch number, is introduced into UNet architecture to alleviate the computation cost at a large extent. On the other hand, the feature misalignment is taken into consideration and a river-like Feature Alignment Flow is proposed to generate spatial deviation and correct the features. The architecture achieves strongly competitive results on public Synapse and DRIVE dataset with pretty light model size and computation requirement. The results show that this is a pretty promising network for future deployment in reality. |
|
| Keywords:computer technology; vision transformer; medical image semantic segmentation; feature alignment; lightweight |
|
