Efficient brain tumor segmentation using Swin transformer and enhanced local self-attention

Fully convolutional neural networks architectures have proven to be useful for brain tumor segmentation tasks. However, their performance in learning long-range dependencies is limited to their localized receptive fields. On the other hand, vision transformers (ViTs), essentially based on a multi-head self-attention mechanism, which generates attention maps to aggregate spatial information dynamically, have outperformed convolutional neural networks (CNNs). Inspired by the recent success of ViT models for the medical images segmentation, we propose in this paper a new network based on Swin transformer for semantic brain tumor segmentation.

The proposed method for brain tumor segmentation combines Transformer and CNN modules as an encoder-decoder structure. The encoder incorporates ELSA transformer blocks used to enhance local detailed feature extraction. The extracted feature representations are fed to the decoder part via skip connections. The encoder part includes channel squeeze and spatial excitation blocks, which enable the extracted features to be more informative both spatially and channel-wise.

The method is evaluated on the public BraTS 2021 datasets containing 1251 cases of brain images, each with four 3D MRI modalities. Our proposed approach achieved excellent segmentation results with an average Dice score of 89.77% and an average Hausdorff distance of 8.90 mm.

We developed an automated framework for brain tumor segmentation using Swin transformer and enhanced local self-attention. Experimental results show that our method outperforms state-of-th-art 3D algorithms for brain tumor segmentation.