MAFDE-Net: Multipath Attention-Fusion-Based Dual-Encoder Network for Undersampled MRI Segmentation

Magnetic Resonance Imaging (MRI) plays a crucial role in medical diagnosis, but previous studies have mainly relied on fully-sampled magnitude images for segmentation. However, prolonged k-space acquisition may cause discomfort and motion artifacts in patients, and undersampling techniques are commonly used to address these issues. Conventional methods often adopt a strategy of first reconstructing and then segmenting, but this approach neglects the influence of reconstruction on the downstream segmentation task. In view of this, integrating undersampled MRI reconstruction with segmentation and improving undersampled segmentation performance via joint training has emerged as a promising strategy. Therefore, we propose a novel network, MAFDE-Net, that integrates reconstruction and segmentation into a unified framework. The network enhances undersampled MRI segmentation performance through a joint learning mechanism. The proposed framework integrates the R2N branch containing four Res2Net modules, the CNN-Transformer (CT) branch, and the Multipath Attention-Fusion (MAF) module synergistically combining features from both branches. In addition, we include an Inverted Residual (IR) module in the decoder stage to effectively integrate features extracted during the encoding stage. The Dynamic Upsampling (DU) module is introduced to enhance the final upsampling quality. Simulation experiments show that the undersampled segmentation performance of MAFDE-Net on three datasets significantly outperforms the joint model (RecSeg) and seven baseline models (considering reconstruction and segmentation as serial tasks). Additionally, the joint learning mechanism adopted by MAFDE-Net is not limited to undersampling scenarios; it also outperforms single-task models in fully-sampled MRI segmentation tasks, expanding its application scenarios and potential impact.