Fast ungated five-dimensional cardiac MRI on a 1.5 T MR-linac for MRI-guided radiotherapy - public k-space dataset

Description This repository contains the public k-space dataset, reconstructed motion states, validation data, and example code accompanying the publication "Fast ungated five-dimensional cardiac MRI on a 1.5 T MR-linac for MRI-guided radiotherapy" The dataset consists of 10 distinct patient/volunteer files formatted as HDF5 (.h5), plus accompanying example code demonstrating how to load and parse the high-dimensional data arrays. Abstract Background: Stereotactic arrhythmia radio-ablation (STAR) for patients with ventricular tachycardia is currently limited by complex cardiorespiratory motion. Hence, patient-specific five-dimensional magnetic resonance imaging (5D-MRI) motion models are required to overcome this clinical challenge. However, current 5D-MRI methods are hampered by long acquisition and reconstruction times, limiting their clinical availability for STAR treatments. Objective: To develop and validate a fast, ungated 5D-MRI reconstruction method, providing personalized motion characterization to support MRI-guided STAR treatments. Methods: We propose a fast, ungated 5D-MRI reconstruction method based on the previously established cardiac MR-MOTUS framework. The method utilizes a variable-density 3D Cartesian acquisition with a joint optimization framework to reconstruct a motion-corrected reference image and low-rank deformation vector fields (DVFs). By exploiting the low rank structure of the DVFs, we explicitly disentangle respiratory and cardiac motion during optimization. After optimization, the DVFs are used as navigators for cardiorespiratory motion, enabling fast 5D-MRI reconstruction with a retrospectively adjustable number of motion states. Validation was performed using digital and physical cardiorespiratory phantoms. Furthermore, the approach was evaluated in vivo using 10 healthy volunteers, comparing motion consistency with standard 2D cine MRI. Results: Validation of 5D CMR-MOTUS using digital and physical phantoms demonstrated accurate 5D-MRI reconstruction. In the physical phantom, 5D CMR-MOTUS achieved a left-ventricle DICE coefficient of 0.96 ± 0.01. In the volunteer cohort, the 5D-MRI scans showed strong motion consistency when compared to 2D cine MRI, yielding a cardiac motion magnitude error of 0.1 ± 0.9 mm and a respiratory motion magnitude error of 0.2 ± 2.9 mm. Crucially, 5D-MRI data were acquired in one minute and reconstructed in six minutes. Conclusions: The proposed 5D-MRI method enables rapid, high-quality, and personalized motion characterization, demonstrating potential for integration into MRI-guided STAR treatments. Dataset Structure Each of the 10 volunteer/subject data files follows an identical hierarchical HDF5 layout divided into a 2D validation arm and a 5D reconstruction arm: └── exam/ ├── 2D/ # Reference 2D validation data │ ├── Acquisition parameters/ # Sequence metadata (TE/TR, etc.) │ ├── original_cine (T, X, Y) # [float32] Time-resolved validation cine │ ├── reference_image (X, Y) # [complex64] Stationary reference frame │ ├── reference_segmentation (X, Y) # [int16] Manually-segmented reference image │ ├── segmentation_structures (7,) # Struct array mapping Label IDs to string names │ ├── spatial_spline_basis (rank, dim, L, L) # [float32] Spatial DVF representation using a cubic B-spline basis │ └── temporal_basis (rank, 1, T) # [float32] Reconstructed temporal respiratory/cardiac basis └── 5D/ # Raw and reconstructed 5D dataset ├── Acquisition parameters/ # Sequence metadata (TE/TR, etc.) ├── espirit_csm (coil, X, Y, Z) # [complex64] ESPIRiT coil sensitivity maps (8 channels) ├── kspace (readout, T, phase_encodes_per_shot, coil) # [complex64] Raw undersampled k-space data ├── recon_10_10 (X, Y, Z, 10, 10) # [complex64] Reconstructed 5D-MRI (10 resp x 10 cardiac phases) ├── recon_20_20 (X, Y, Z, 20, 20) # [complex64] Reconstructed 5D-MRI (20 resp x 20 cardiac phases) ├── reference_image (X, Y, Z) # [complex64] Motion-corrected 3D reference volume ├── reference_segmentation (X, Y, Z) # [int16] Manually segmented reference image ├── segmentation_structures (8,) # Struct array mapping Label IDs to string names ├── spatial_spline_basis (rank, 3, L, L, L) # [float32] 3D low-rank spatial basis using a cubic B-spline basis ├── temporal_basis (rank, 1, T) # [float32] low-rank temporal motion paths ├── y_phase_encodes (T, phase_encodes_per_shot) # [int32] Phase encode indices along the Y-axis └── z_phase_encodes (T, phase_encodes_per_shot) # [int32] Phase encode indices along the Z-axis Segmentation Labels Map Manual segmentations mapped within segmentation_structures represent the following anatomical structures delineated using nnInteractive: 2D Cine: Left ventricle, right ventricle, myocardium, liver, left lung, and right lung. 5D Volumes: Left ventricle, right ventricle, myocardium, liver, aorta, left lung, and right lung. Sequence Parameters Raw k-space data were acquired on a 1.5 T Elekta Unity MR-linac using an 8-channel phased-array coil setup. The data were acquired without gating, triggering, or use of external navigators using the following acquisition parameters: Parameter 2D Cine MRI 5D-MRI (OPRA bSSFP) Pulse Sequence Balanced gradient echo Balanced steady-state free precession Readout Direction FH RL Number of Coils 8 8 Slice Thickness (mm) 9 - TR / TE (ms) 3.2 / 1.6 3.8 / 1.9 Readouts per Frame 38 26 Temporal Resolution 8.3 Hz 10.1 Hz Flip Angle 50° 50° Bandwidth (Hz/px) 1102 1293 FOV (mm) 288 × 256 248 × 200 × 385 Spatial Resolution (mm) 1.6 × 1.6 2 × 2 × 2 Matrix Size 180 × 160 124 × 100 × 193 Scan Time 10 s 60 s Usage & Getting Started An example Jupyter Notebook is bundled with this repository to assist with quick loading and visualization of the multidimensional matrices using h5py and numpy. Alongside this Notebook are the B-spline primitive operations based on the TorchIR framework [1,2], and the utilities to bin a signal into N motion states. Ethics and Data Privacy Informed consent was obtained from each subject prior to imaging in accordance with institutional guidelines and Local Ethics Committee approval (NL59820.041.7). All data elements have been structurally de-identified to comply with privacy regulations. References [1] Bob D. de Vos, Floris F. Berendsen, Max A. Viergever, Marius Staring and Ivana Išgum, "End-to-end unsupervised deformable image registration with a convolutional neural network," in Deep learning in medical image analysis and multimodal learning for clinical decision support. Springer, Cham, 2017. p. 204-212, doi: 10.1007/978-3-319-67558-9_24 https://link.springer.com/chapter/10.1007%2F978-3-319-67558-9_24 [2] Bob D. de Vos, Floris F. Berendsen, Max A. Viergever, Hessam Sokooti, Marius Staring and Ivana Išgum "A deep learning framework for unsupervised affine and deformable image registration," Medical image analysis, vol. 52, pp. 128-143, Feb. 2019, doi: 10.1016/j.media.2018.11.010 https://www.sciencedirect.com/science/article/pii/S1361841518300495