Feature Learning by Attention and Ensemble with 3D U-Net to Glioma Tumor Segmentation

BraTS2021 Task1 is research on segmentation of intrinsically heterogeneous brain glioblastoma sub-regions in mpMRI scans. Base on BraTS 2020 top ten team's solution (open brats2020, ranked among the top ten teams work), we proposed a similar as 3D U-Net neural network, called as TE U-Net, to differentiate glioma sub-regions class. According that automatically learns to focus on sub-regions class structures of varying shapes and sizes, we proposed TE U-Net which is similar with U-Net++ network architecture. Firstly, we reserved encoder second and third stage's skip connect design, then also cut off first stage skip connect design. Secondly, multiple stage features through attention gate block before features skip connect, so as to ensemble channels and space region information to suppress irrelevant regions. Finally, in order to improve model performance, on network post-processing stage, we ensemble multiple similar 3D U-Net with attention module. On the online validation database, the TE U-Net architecture get best result is that the GD-enhancing tumor (ET) dice is 83.79%, the peri-tumoral edematous/invaded tissue (TC) dice is 86.47%, and the necrotic tumor core (WT) dice is 91.98%, Hausdorff (95%) values is 6.39,7.81,3.86 and Sensitivity values is 82.20%, 83.99%, 91.92% respectively. And our solution achieved a dice of 85.62%,86.70%,90.64% for ET,TC and WT, as well as Hausdorff(95%) is 18.70,21.06,10.88 on final private test dataset.