3D ATTENTION U-NET BASED MODEL FOR MULTI-REGION BRAIN TUMOR SEGMENTATION

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.