Mptgan: A Multimodal Prior-Based Triple-Branch Network For Fast Prostate MRI Reconstruction

As the most commonly used radiological examination for prostate disorder diagnosis, Magnetic Resonance Imaging (MRI) acquisition is very time-consuming. To accelerate prostate MRI reconstruction while maintaining high quality, this paper provides MPTGAN, a multimodal prior-based triple-branch network. MPTGAN guides the reconstruction of time-consuming MRI modality by utilizing time-efficient MRI modality as prior knowledge, thus the massive loss of low and high-frequency information in under-sampled data could be replenished. In particular, a hybrid feature interaction module (HFIM) is employed to integrate and fuse the latent relation clues between multimodal MRIs. Besides, we also design a triple-attention upsampling module (TAUM) to capture the salient features from the fused multimodal information. We collected 7300 pairs of T1 and T2 images from 387 prostate disorder patients and applied MPTGAN on them. Experiments show that MPTGAN significantly improved the reconstruction quality of prostate MRI and outperformed state-of-the-art methods.