Groupwise Registration with Physics-Informed Test-Time Adaptation on Multi-parametric Cardiac MRI

TL;DR

Multiparametric CMR maps such as $T_1$ and $T_2$ are prone to misalignment from motion, complicating pixel-wise analysis. The authors introduce physics-informed test-time adaptation (PI-TTA) with a hierarchical two-level groupwise registration to enable robust cross-contrast alignment by generating inversion-recovery–like synthetic references for case-specific fine-tuning. The method pre-trains on model-agnostic data and adapts per test series, replacing the similarity term with $L_2$ when matching synthetic references. Results on healthy volunteers show improved myocardial boundary alignment and higher $R^2$ fits and Dice scores between T1 and T2 maps, with runtime around $10$ seconds, suggesting potential for real-time clinical use. The approach is flexible to other multiparametric mappings and multiple physics models.

Abstract

Multiparametric mapping MRI has become a viable tool for myocardial tissue characterization. However, misalignment between multiparametric maps makes pixel-wise analysis challenging. To address this challenge, we developed a generalizable physics-informed deep-learning model using test-time adaptation to enable group image registration across contrast weighted images acquired from multiple physical models (e.g., a T1 mapping model and T2 mapping model). The physics-informed adaptation utilized the synthetic images from specific physics model as registration reference, allows for transductive learning for various tissue contrast. We validated the model in healthy volunteers with various MRI sequences, demonstrating its improvement for multi-modal registration with a wide range of image contrast variability.

Figures (4)

• Figure 1: Overview of the physics-informed test-time adaptation. The model is pre-trained using training data, and the pre-trained model is adapted to obtain desired capabilities on specific test images according to different physics models. The similarity term is changed from normalized mutual information to mean squared error during fine-tuning. In this figure, we use the data following the $T_1$-relaxation model as an example. In practice, it can be replaced with others based on the data modality. The rPCA-GroupRegNet is built based on the previous work li2023contrast
• Figure 2: Hierarchical two-level 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 intra-sequence in the second level. (B) The physics model, T1+(1) model, applied in the second level for inter-subject registration are shown.
• Figure 3: Representative figures and boxplots compare the performance of three methods on T1-weighted and T2-weighted data, using the scanner MOCO, pca-relax model and our proposed method. Our proposed approach (w/ PI) shows consistent improvement comparing to other approaches, especially around the myocardium boundary as indicated by the white arrows.
• Figure 4: Representative figures and boxplots compare the performance between single-level and second-level registration. The representative figure show the improvement of the alignment between T1 and T2 maps using 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 (p < 0.05) in dice score using second-level registration.