MR-Physics-Guided Deep Learning for Robust Multiparametric CMR Motion Correction

Presenting Author(s)

• XL

  Xinqi Li, MSc

  Ph.D student
  Cedars-Sinai Medical Center
  Berlin, Berlin, Germany

Primary Author(s)

• XL

  Xinqi Li, MSc

  Ph.D student
  Cedars-Sinai Medical Center
  Berlin, Berlin, Germany

Co-Author(s)

• YZ

  Yi Zhang

  PhD candidate
  Delft University of Technology, Netherlands
• 

  Li-Ting Huang, MD, MSc

  radiologist
  National Cheng Kung University Hospital
  Los Angeles, California, United States
• 

  Thoralf Niendorf, PhD

  Prof.Dr.
  Berlin Ultrahigh Field Facility (B.U.F.F.), Max Delbrueck Center for Molecular Medicine in the Helmholtz Association, Berlin, Germany. Experimental and Clinical Research Center, Charite Medical Faculty and the Max Delbrueck Center for Molecular Medicine in the Helmholtz Association, Berlin, Germany. DZHK (German Centre for Cardiovascular Research), partner site Berlin, Berlin, Germany.
  Berlin, Berlin, Germany
• 

  Kim-Lien Nguyen, MD

  Associate Professor of Cardiovascular Medicine and Radiology
  David Geffen School of Medicine at UCLA and VA Greater Los Angeles Healthcare System
  Los Angeles, California, United States
• MK

  Min-Chi Ku, PhD

  Senior researcher
  Berlin Ultrahigh Field Facility (B.U.F.F.), Max Delbrueck Center for Molecular Medicine in the Helmholtz Association, Berlin, Germany. DZHK (German Centre for Cardiovascular Research), partner site Berlin, Berlin, Germany
  Berlin, Berlin, Germany
• QT

  Qian Tao, PhD

  Assistant Professor
  Delft University of Technology
  Delft, Zuid-Holland, Netherlands
• 

  Hsin-Jung (Randy) Yang, PhD

  Associate Professor
  Biomedical Imaging Research Institute, Cedars-Sinai Medical Center, Los Angeles, CA, USA
  Los Angeles, California, United States

Background: Multiparametric quantitative CMR techniques, such as T1 and T2 mapping, are powerful for characterizing cardiovascular pathology. However, cardiac and respiratory motion can compromise spatial alignment and affect the accuracy of quantitative fitting and multiparametric analysis. Conventional iterative registration methods can correct motion within a single contrast model but often fail across different contrasts. Recent deep learning methods have explored contrast-agnostic registration but remain limited by training domain generalizability.
We propose a physics-informed test-time adaptation (PI-TTA) framework for multiparametric cardiac registration. This approach pre-trains the neural network using a model-agnostic strategy¹ and incorporates MR signal evolution models, enabling adaptation to the specific contrast of each data. This hierarchical framework corrects intra-sequence motion and subsequently aligns multi-contrast acquisitions, facilitating robust registration across diverse quantitative CMR mappings.

Methods: The proposed hierarchical two-level registration pipeline (Fig.1A) operates as follows:

1. Intra-sequence correction: Each image series is first registered independently using PI-TTA, which leverages physics-based contrast models during test-time fine-tuning. For example, in inversion recovery sequences, synthetic series generated from spin-recovery models guided fine-tuning (Fig.1B).


2. Inter-sequence co-registration: Motion-corrected series from the first level were subsequently aligned across contrasts. For T1–T2 co-registration, the first T2 image was treated as a T1-weighted surrogate to bridge contrast differences (Fig.3A).

The model was pre-trained on 120 post-contrast MOLLI images and tested on three datasets: a) in-domain post-contrast MOLLI (n=15), b) out-of-domain large-animal MOLLI with anatomical differences (n=7), and c) T2-weighted images from healthy volunteers (n=10). Comparisons were made with Siemens MOCO² and a prior deep-learning method (PCA-Relax³). The registration quality is assessed with a groupwise dice score and the R2 of the fitted maps⁴.

Results: PI-TTA demonstrated improved registration quality across domains, particularly at the endocardial boundary (Fig.2). Quantitatively, dice scores (improvement compared to the original images) and fitting R² values improved significantly versus both MOCO and PCA-Relax (P< 0.05). The second-level inter-sequence registration further enhanced alignment of T1 and T2 maps, improving overlap metrics and demonstrating robust multiparametric integration (Fig.3).

Conclusion: The proposed PI-TTA framework enables accurate, contrast-robust motion correction and multiparametric registration by integrating MR-physics-based test-time adaptation with hierarchical alignment. This approach improves the fidelity of quantitative mapping and facilitates pixel-wise multiparametric CMR analysis. It broadly applies to diverse mapping protocols and can potentially strengthen clinical utility of quantitative CMR.

Figure 1 Overview of the hierarchical registration pipeline. (A) The generalizable two-level registration pipeline is shown, where the moving volume is first registered within the sequences in the first level and then registered inter-sequence in the second level. (B) The detailed network architecture with physics-informed test-time adaptation (PI-TTA), using the T1 relaxation model as an example. (C) The experimental setup used to compare the performance and evaluate the generalizability of our proposed method.
Figure 2: First-level intra-series registration performance. (A) The registration results of in-domain T1 MOLLI sequence(Human, post contrast), (B) Anatomical difference (from human to large animal), and (C) Model difference (from T1 to T2 Relaxation). The representative images and boxplot demonstrated the improvement of using our proposed method comparing to others in both global dice score and fitting performance in all three cases.
Figure 3 Second-level inter-series registration performance. (A) The T1+(1) physics model was applied in the second level for inter-subject registration between T1 and T2 weighted images. This approach treats the first T2-weighted image as a fully recovered T1-weighted image. The images were preprocessed using histogram matching to normalize the intensity difference between the imaging protocols. The combined series, including all T1-weighted images and the intensity-matched first T2-weighted image, then serves as input to the second-level registration. (B) Representative images and dice score comparison of T1 and T2 maps before and after the second level registration. The red and green masks denote the T1 and T2 mapping's myocardial masks and the white region denote the overlay region. The statistical results showed significant improvement in the dice score after registration.