Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DuDoUniNeXt: Dual-domain unified hybrid model for single and multi-contrast undersampled MRI reconstruction

Ziqi Gao, Yue Zhang, Xinwen Liu, Kaiyan Li, S. Kevin Zhou

TL;DR

Experimental results demonstrate that the proposed DuDoUniNeXt model surpasses state-of-the-art SC and MC models significantly, and the effectiveness of the proposed hybrid backbone, AdaC2F, PaSS, and the dual-domain unified learning scheme.

Abstract

Multi-contrast (MC) Magnetic Resonance Imaging (MRI) reconstruction aims to incorporate a reference image of auxiliary modality to guide the reconstruction process of the target modality. Known MC reconstruction methods perform well with a fully sampled reference image, but usually exhibit inferior performance, compared to single-contrast (SC) methods, when the reference image is missing or of low quality. To address this issue, we propose DuDoUniNeXt, a unified dual-domain MRI reconstruction network that can accommodate to scenarios involving absent, low-quality, and high-quality reference images. DuDoUniNeXt adopts a hybrid backbone that combines CNN and ViT, enabling specific adjustment of image domain and k-space reconstruction. Specifically, an adaptive coarse-to-fine feature fusion module (AdaC2F) is devised to dynamically process the information from reference images of varying qualities. Besides, a partially shared shallow feature extractor (PaSS) is proposed, which uses shared and distinct parameters to handle consistent and discrepancy information among contrasts. Experimental results demonstrate that the proposed model surpasses state-of-the-art SC and MC models significantly. Ablation studies show the effectiveness of the proposed hybrid backbone, AdaC2F, PaSS, and the dual-domain unified learning scheme.

DuDoUniNeXt: Dual-domain unified hybrid model for single and multi-contrast undersampled MRI reconstruction

TL;DR

Experimental results demonstrate that the proposed DuDoUniNeXt model surpasses state-of-the-art SC and MC models significantly, and the effectiveness of the proposed hybrid backbone, AdaC2F, PaSS, and the dual-domain unified learning scheme.

Abstract

Multi-contrast (MC) Magnetic Resonance Imaging (MRI) reconstruction aims to incorporate a reference image of auxiliary modality to guide the reconstruction process of the target modality. Known MC reconstruction methods perform well with a fully sampled reference image, but usually exhibit inferior performance, compared to single-contrast (SC) methods, when the reference image is missing or of low quality. To address this issue, we propose DuDoUniNeXt, a unified dual-domain MRI reconstruction network that can accommodate to scenarios involving absent, low-quality, and high-quality reference images. DuDoUniNeXt adopts a hybrid backbone that combines CNN and ViT, enabling specific adjustment of image domain and k-space reconstruction. Specifically, an adaptive coarse-to-fine feature fusion module (AdaC2F) is devised to dynamically process the information from reference images of varying qualities. Besides, a partially shared shallow feature extractor (PaSS) is proposed, which uses shared and distinct parameters to handle consistent and discrepancy information among contrasts. Experimental results demonstrate that the proposed model surpasses state-of-the-art SC and MC models significantly. Ablation studies show the effectiveness of the proposed hybrid backbone, AdaC2F, PaSS, and the dual-domain unified learning scheme.
Paper Structure (11 sections, 3 equations, 4 figures, 2 tables)

This paper contains 11 sections, 3 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Dual-domain Unified Learning Framework
  4. Contrast-aware Dynamic Encoder
  5. Domain-specific CNN-ViT hybrid backbone
  6. Training scheme
  7. Experiment
  8. Setup
  9. Main results
  10. Ablation studies
  11. Conclusion

Figures (4)

  • Figure 1: The main idea of DuDoUniNeXt. (a) The reconstruction results of different models with absent, an LQ, or an HQ auxiliary image. The upper part shows the results of the MC model, the middle part shows the results of the SC model, and the bottom is the results of the proposed DuDoUniNeXt. The PSNR values (dB) of the reconstructed images are shown in the upper right corners. (b) The reconstruction emphasis of CNN, ViT, and our hybrid backbone. The green and red boxes highlight two typical patterns: highly structured lines and isolated fine details.
  • Figure 2: The overall dual-domain unified learning framework. Each recurrent block contains one network for single-contrast k-space restoration, K-NeXt, and one network for unified image restoration, I-UniNeXt, with two interleaved DCs. Availability condition (AC) keeps I-UniNeXt notified for different conditions of $I_{ref}$.
  • Figure 3: The architecture of I-UniNeXt. Panel (a) shows the overall architecture of I-UniNeXt, with the detailed architecture of the PaSS illustrated in the blue box and XBB illustrated in the orange box. (b) and (c) show the dedicated structure of AdaC2F and the domain-specific Hybrid Block (XBB$_j$), respectively.
  • Figure 4: Qualitative comparisons of 5 $\times$ undersampled T2 and PD reconstruction under different $I_{ref}$ conditions. Corresponding error maps are illustrated in BWR colormaps.