Abstract:Recently, U-Net architecture with its strong adaptability has become prevalent in the field of MRI brain tumor segmentation. Meanwhile, researchers have demonstrated that introducing attention mechanisms, especially self-attention, into U-Net can effectively improve the performance of segmentation models. However, the self-attention has disadvantages of heavy computational burden, quadratic complexity as well as ignoring the potential correlations between different samples. Besides, current attention segmentation models seldom focus on adaptively computing the receptive field of tumor images that may capture discriminant information effectively. To address these issues, we propose a novel 3D U-Net related brain tumor segmentation model dubbed as self-calibrated attention U-Net (SCAU-Net) in this work, which simultaneously introduces two lightweight modules, i.e., external attention module and self-calibrated convolution module, into a single U-Net. More specifically, SCAU-Net embeds the external attention into the skip connection to better utilize encoding features for semantic up-sampling, and it leverages several 3D self-calibrated convolution modules to replace the original convolution layers, which adaptively computes the receptive field of tumor images for effective segmentation. SCAU-Net achieves segmentation results on the BraTS 2020 validation dataset with the dice similarity coefficient of 0.905, 0.821 and 0.781 and the 95% Hausdorff distance (HD95) of 4.0, 9.7 and 29.3 on the whole tumor, tumor core and enhancing tumor, respectively. Similarly, competitive results are obtained on BraTS 2018 and BraTS 2019 validation datasets. Experimental results demonstrate that SCAU-Net outperforms its baseline and achieves outstanding performance compared to various representative brain tumor models.

SCAU-net: 3D self-calibrated attention U-Net for brain tumor segmentation