Core Concepts
The MBDRes-U-Net model achieves accurate brain tumor segmentation from multimodal MRI images while significantly reducing computational complexity compared to traditional 3D U-Net models.
Stats
Tumor regions account for only 1.5% of the MRI images in the BraTS dataset.
Enhanced tumors (ETs) account for only 11% of the whole tumor (WT) images in the BraTS dataset.
The MBDRes-U-Net model reduces the parameters of the traditional 3D U-Net model by a factor of four.
The computational complexity is reduced by 1643.75 G compared to the 3D U-Net model.
The MBDRes-U-Net model achieves a 3.2%, 1.8%, and 13.6% improvement in Dice scores for ET, WT, and TC segmentation, respectively, compared to the 3D U-Net model on the BraTS 2018 dataset.