Mamba-Based Modality Disentanglement Network for Multi-Contrast MRI Reconstruction

TL;DR

This work tackles the challenge of long MRI acquisition times by introducing MambaMDN, a dual-domain framework that first completes undersampled target K-space using a fully sampled reference, producing a structurally preserved yet modality-mixed input. It then employs a Mamba-based modality disentanglement network with progressive refinement to purge reference-specific features, yielding accurate target reconstructions. Across IXI, BraTS, and M4raw datasets, MambaMDN achieves state-of-the-art PSNR/SSIM with favorable computational efficiency compared to Transformer-based baselines, under both ×4 and ×8 undersampling. The approach holds promise for clinical deployment in resource-limited settings by delivering high-quality multi-contrast MRI from highly undersampled data while mitigating cross-modal interference.

Abstract

Magnetic resonance imaging (MRI) is a cornerstone of modern clinical diagnosis, offering unparalleled soft-tissue contrast without ionizing radiation. However, prolonged scan times remain a major barrier to patient throughput and comfort. Existing accelerated MRI techniques often struggle with two key challenges: (1) failure to effectively utilize inherent K-space prior information, leading to persistent aliasing artifacts from zero-filled inputs; and (2) contamination of target reconstruction quality by irrelevant information when employing multi-contrast fusion strategies. To overcome these challenges, we present MambaMDN, a dual-domain framework for multi-contrast MRI reconstruction. Our approach first employs fully-sampled reference K-space data to complete the undersampled target data, generating structurally aligned but modality-mixed inputs. Subsequently, we develop a Mamba-based modality disentanglement network to extract and remove reference-specific features from the mixed representation. Furthermore, we introduce an iterative refinement mechanism to progressively enhance reconstruction accuracy through repeated feature purification. Extensive experiments demonstrate that MambaMDN can significantly outperform existing multi-contrast reconstruction methods.

Figures (11)

• Figure 1: Comparison of multi-contrast MRI reconstruction strategies. (a) Fusion-based approaches fuse the undersampled target image $I_{tar}^{us}$ with the high-quality reference image $I_{ref}$, which may introduce irrelevant features. (b) Our proposed disentanglement-based method first complements the undersampled target K-space using the reference modality, generating a structure-preserved but modality-mixed image $I_{mix}^{fs}$. A modality disentanglement network then removes reference-specific features to yield a clean target reconstruction $I_{tar}$.
• Figure 2: Visualization of modality-mixed images generated under different accelerations. As the undersampling factor increases, more information is borrowed from the reference modality, resulting in a hybrid reconstruction that exhibits enhanced PD-like contrast (e.g., higher gray matter intensity and reduced T2-specific fluid signal).
• Figure 3: Overview of the proposed MambaMDN framework. The undersampled target K-space data $K_{tar}^{us}$ is complemented by the fully sampled reference data $K_{ref}^{fs}$ to generate a contrast-mixed image $I_{mix}$. Furthermore, we use two encoders to extract high-level features $F_{mix}$ and $F_{ref}$, which are progressively disentangled by Mamba-based modules to reconstruct the target image. Finally, a DC layer is employed to ensure data consistency to get the final result.
• Figure 4: The proposed KCM module. The fully-sampled reference (PD) image is transformed to the K-space domain and used to fill the missing regions of the undersampled target (T2) K-space, yielding an artifact-free but modality-mixed image for the reconstruction network.
• Figure 5: Illustration of the proposed Mamba-based Modality Disentanglement (MMD) module. (A) The overall structure of MMD disentangles modality-specific features from the mixed representation using reference modulation. (B) The original Mamba block. Inspired by it, we designed our MMD module. (C) The SS2D block proposed by Vmamba liu2024vmamba enables the MMD to effectively model spatially long-range features.