LFBTS: Enhanced Multimodality MRI Fusion for Brain Tumor Segmentation With Limited Computational Resources

Abstract

Efficient and accurate segmentation of brain tumors in multimodal magnetic resonance imaging (MRI) is crucial for clinical diagnosis and treatment planning. Traditional methods tend to concentrate solely on feature extraction from individual modalities, overlooking the substantial potential of multimodal feature fusion in enhancing segmentation performance. In this paper, we present a novel method that not only integrates salient features from different modalities strategically but also takes into account the constraints imposed by limited computational resources, ensuring both accuracy and efficiency. Two key modules, the attention-guided cross-modality fusion module (ACFM) and the hierarchical asymmetric convolution module (HACM), were designed to leverage the distinct modalities and the varying information focuses found within different dimensions. The ACFM is based on a transformer framework, utilizing self-attention and cross-attention mechanisms. These mechanisms enable the capture of both local and global dependencies within and between different MRI modalities. This design allows for the effective fusion of complementary features from multiple modalities, thereby enhancing segmentation performance by leveraging the valuable information contained in each modality. Meanwhile, the HACM reduces computational complexity using a pseudo-3D convolution approach. This approach breaks down 3D convolutions into components along the transverse and sagittal axes. Unlike traditional 2D convolutions, this method preserves essential spatial information across dimensions. It ensures accurate segmentation while maximizing efficiency by capitalizing on the varying focus of information in different spatial planes. This approach takes advantage of the varying information density in these dimensions, achieving a balance between accuracy and efficiency. Through extensive experiments on the BraTS2021 dataset, our proposed modality fusion-based network under limited resources (LFBTS) achieves dice scores of 0.925, 0.911, and 0.886 for whole tumor (WT), tumor core (TC), and enhanced tumor (ET), respectively. These results outperform state-of-the-art (SOTA) models and consistently demonstrate superiority over models developed in the preceding 2 years. This highlights the potential of our approach in advancing brain tumor segmentation and improving clinical decision-making, particularly in settings with limited resources.