ARS-Net:3D Multimodal Feature Recalibration Fusion Network for Brain Tumor Image Segmentation

To enhance the diagnosis and treatment of brain tumors, precise segmentation of the entire tumor region is crucial. Medical image fusion is a method that integrates complementary information from different modal images into a synthetically enriched image. In current research, 2D slice fusion methods are prevalent, but they often lose essential spatial context information when fusing volumetric structures in medical images. Furthermore, existing 3D medical image fusion techniques fail to fully exploit the features of the source modality, leading to the loss of crucial modality information. Addressing these challenges, this paper proposes an Average Recalibration Fusion Segmentation Network (ARS-Net) for the task of 3D multi-modal brain tumor segmentation. The network comprises two-stage feature fusion: a fusion module based on simple averaging and a feature recalibration module aimed at maximizing the utilization of learned neural network features. By appropriately integrating spatial information from the features, we can maintain the relative priority of activated regions, thereby facilitating channel recalibration. Additionally, we design a novel loss function specifically considering the features of different MRI modalities to better preserve modality information. In a series of experiments, our method demonstrates more competitive performance in both visual quality and objective evaluation compared to other representative 3D and 2D medical image fusion methods. Further experiments validate the effectiveness of the proposed method in brain tumor segmentation, showing a significant improvement in segmentation accuracy.