3D BEFUNET MODEL FOR BRAIN TUMOR SEGMENTATION

Abstract

Brain cancer is a devastating and potentially fatal illness that has a profoundly detrimental impact on its victims' 
lives. However, early brain tumor detection is a difficult task and an unmet need. This work developed models for 
reliable and clinically applicable brain tumor analysis by extending BEFUnet to 3D processing and optimizing it for 
multi-class segmentation. A comparative evaluation was conducted to assess the segmentation performance of the 
3D BEFUnet against four benchmark models which are nnU-Net, 3DUV-NetR+, Scale Attention Network, and 
TMA-TransBTS using the BraTS 2020 dataset. The performance was quantified using the Dice Similarity 
Coefficient (DSC) across three tumor subregions: Enhancing Tumor (ET), Tumor Core (TC), and Whole Tumor 
(WT). The results reveal that while the model achieved moderate Dice coefficients (ET: 0.331, TC: 0.546, WT: 
0.605), it significantly outperformed benchmark models in terms of boundary accuracy (HD95) and specificity for 
the Tumor Core (TC) and Whole Tumor (WT) regions. These findings demonstrate that 3D BEFUnet achieves a 
strong balance between computational efficiency and segmentation precision which is an essential trade-off for real
world deployment in clinical environments with limited computational resources.