Enhancing Brain Tumor Segmentation with Deep Supervision and Attention Mechanisms: Advances in the nnU-Net Framework

Yi-Zeng Fang, Juinn-Dar Huang

Abstract: In the evolving landscape of medical imaging, achieving precise brain tumor segmentation from MRI images is critical for diagnosis and treatment. Our work, presented at BraTS Challenge 2024, advances the nnU-Net framework by incorporating deep supervision and the Convolutional Block Attention Module (CBAM), enhancing model performance through focused feature learning and attention mechanisms. Deep supervision ensures effective feature extraction by reinforcing gradient flow, while CBAM optimizes the model focus on relevant features, significantly improving segmentation accuracy. Tested on the BraTS 2024 challenge validation dataset, our approach yielded Lesion-Wise Dice of 0.765, 0.823, and 0.797, and Lesion-Wise Hausdorff Distance of 58.667, 35.485, and 51.930 for whole tumor (WT), tumor core (TC), and enhancing tumor (ET) regions, respectively. These results highlight the benefits of our method, offering a notable contribution to the field of neurooncology diagnostics.