Abstract:
Brain tumors are characterized by the abnormal growth of cells within or near the brain
tissues. Accurate segmentation of brain tumors is crucial for effective clinical decision mak-
ing. The traditional models encounter challenges in accurately delineating tumor regions. In
addition, training robust segmentation models on high-resolution magnetic resonance data
requires high computational resources. This work proposes 3D attention-based U-Net archi-
tecture for multi-region segmentation of brain tumors using a single stacked multi-modal vol-
ume. Incorporating the Squeeze and Excitation (SE) attention mechanism into the encoder
side of the proposed model facilitates the capture of fine-grained details and the prioritiza-
tion of significant regions through segmentation areas. This work uses a publicly available
Brats2020 (brain tumor segmentation) dataset. The segmentation performance of the model
is optimized by evaluation matrices such as Dice coefficient metrics. The proposed model
achieved Dice coefficient scores of 0.84, 0.89, and 0.86 for the Enhancing Tumor, the Tumor
Core, and the entire Tumor, respectively. This work contributes to accurate brain tumor
identification and highlights the transformative role of deep learning in advancing medical
image analysis for improved patient care.
