<![CDATA[Segmentation of brain tumors in volumetric medical images is challenging due to the complexities of the tumor structure, the types, and the heavy-weight 3D data processing. In contrast, 2D-based segmentation methods on the slice data reduce the amount of information due to the anisotropic shape of the tumors and lead to poor segmentation results. This study proposes a 3D network structure combining ReSidual U-Block (RSU), custom dilated block, and U2-Net for automatic segmentation of brain tumors from MRI images, namely 3D RSU U2-Net+. The RSU and custom dilated block are embedded and joined in the nested U-Net structure to obtain multi-resolution features and global information, enhancing segmentation accuracy while reducing computational overhead. The proposed method outperformed the segmentation results of the standard U-Net, on brain tumor data in the medical segmentation Decathlon (MSD) dataset. The proposed model achieves an average validation soft dice loss of 0.1320 and dice score coefficient of 78% and intersection over union of 64% for testing. Although having 3 times parameters, the model requires less GPU time (397.7 minutes) than U-Net (433.6 minutes), demonstrating improved computational efficiency resulting from the effective use of residual and dilated blocks. Moreover, the model achieves 75.4% average sensitivity and 99% specificity for edema, enhancing, and non-enhancing tumors. These experimental results show that the 3D RSU U2-Net+ has been able to outperform the U-Net. However, the model’s performance on non-enhancing tumors remains relatively lower compared to other tumor types, indicating on opportunity for further optimization.]]>
Copyrights © 2026
