FUNet: Frequency Channel Multi-Modal Fusion and Uncertain Region Adjustment Network for brain tumor segmentation

Multi-modal images are crucial for enhancing the performance of brain tumor segmentation. Existing multi-modal brain tumor segmentation methods have the following three main shortcomings: To begin with, framework design remains underexplored in current research. Secondly, effectively fusing multi-modal data, which characterize brain tumors differently, poses a significant challenge. Finally, uncertain and error-prone regions may exist within the fused features, complicating subsequent analysis. To address these issues, we propose Frequency Channel Multi-Modal Fusion and Uncertain Region Adjustment Network (FUNet). FUNet employs a triple-parallel-stream framework to integrate multi-modal information. In the encoder of the multi-modal information learning stream, we design a frequency channel multi-modal fusion module (FCMM), which distinguishes between the complementarity and redundancy of the modal information and mines the intrinsic connection. Additionally, in the decoder, we design an uncertain region adjustment module (URAM), which generates an adjustment factor to enable pixel-wise adjust uncertain error-prone regions existing in the fused features. Experiments on BrsTS 2018 and BraTS-PED 2023 demonstrate that our method achieves better results than other state-of-the-art methods.