DenseTrans: Multimodal Brain Tumor Segmentation Using Swin Transformer

Aiming at the task of automatic brain tumor segmentation, this paper proposes a new DenseTrans network. In order to alleviate problem that convolutional neural networks(CNN) cannot establish long-distance dependence and obtain global context information, swin transformer is introduced into UNet++ network, local feature information extracted by layer in UNet++. then, high resolution layer, shift window operation utilized self-attention learning windows are stacked capability dependency modeling. meanwhile, secondary increase computational complexity caused full transformer, deep separable convolution control layers adopted achieve balance between accuracy segmentation complexity. on BraTs2021 data validation set, model performance as follows: dice dimilarity score was 93.2%,86.2%,88.3% whole tumor,tumor core enhancing tumor, hausdorff distance(95%) values 4.58mm,14.8mm 12.2mm, lightweight with 21.3M parameters 212G flops obtained depth-separable other operations. conclusion, proposed effectively improves tumors has clinical value.