A Multimodal Fusion Generation Network for High-quality MR Image Synthesis
Rui Zhu, Yidan Yan, Jiayao Li, Ruizhi Sun · 2024
Multimodal magnetic resonance (MR) images are important for accurate analysis of the type and extent of lesions. However, obtaining high-quality multimodal MRI images remains challenging due to factors such as equipment and personnel. To solve this problem, we propose a multimodal fusion generation network (MF-Net) for high-quality MR image synthesis, which can synthesize target modality images from two existing modality images. The network has two core modules: The multi-scale Residual Atrous (MRA) module for feature extraction and the Parallel Fusion Attention (PFA) module for attention mechanisms. The MRA module is designed to efficiently extract contextual features at different scales, while the PFA module enhances and fuses features from different modalities effectively. Extensive experiments on BraTS2019 indicate that MF-Net can effectively synthesize target modality images and outperform other single-modality synthesis methods.