An in vivo validation dataset for dynamic volumetric MRI

TL;DR

This work addresses the lack of publicly accessible ground-truth data for validating time-resolved 3D+t MRI reconstructions. It introduces an in vivo dataset where thigh muscles are repeatedly deformed with a pneumatically controlled cuff, enabling both undersampled dynamic data and fully sampled ground-truth validation for one deformation across nine volunteers. The authors provide detailed acquisition protocols (1.5T, multi-coil arrays, CASPR sampling) and demonstrate a binning-based reconstruction approach that aligns undersampled dynamic data with validation images, illustrating the dataset's utility for method development and validation. By including anatomical references, segmentation masks, and accompanying code, the work facilitates rigorous benchmarking and wider adoption of dynamic volumetric MRI techniques.

Abstract

Dynamic volumetric MRI provides valuable information on in vivo motion and biomechanics, with applications spanning cardiac, musculoskeletal, or pulmonary imaging, amongst others. Developing reconstruction methods for time-resolved volumetric MRI is challenging due to the inherently slow acquisition process of MRI, which makes it an active area of research. However, in vivo validation of these methods remains challenging due to the lack of publicly available datasets with fully sampled ground-truth images. Here, we present a publicly available in vivo dataset designed to facilitate the development and validation of dynamic volumetric MRI reconstruction algorithms. Controlled and repeatable deformations of the muscles in the thigh were induced using a pneumatic pressure cuff, enabling the acquisition of both undersampled dynamic data and fully sampled validation images. The dataset comprises multichannel undersampled k-space data from nine healthy volunteers across four different dynamic deformations, with fully sampled validation data for one deformation. Additionally, an anatomical reference scan and muscle segmentation masks are provided for each subject. To illustrate a possible image reconstruction and validation approach, a binning-based reconstruction was performed on the undersampled data from six dynamic repetitions. The resulting images were consistent with the corresponding fully sampled validation images. This dataset offers possibilities for validating and advancing time-resolved volumetric MRI reconstruction methods.

Figures (5)

• Figure 1: Experimental setup, from left to right: pressure gauge, hand pump, long tube (allowing operation outside of the scanner room), and pressure cuff, which is placed around the thigh of the volunteer.
• Figure 2: Pressure profiles for the validation scans and the dynamic scans. Data acquired during the blue-shaded time intervals can be used for dynamic time-resolved reconstructions, while data acquired during the orange-shaded time intervals can serve as validation data for the dynamic reconstructions. Dynamic scans 2 and 4 were acquired during an isometric knee flexion task. During dynamic scans 3 and 4 the pressure cuff was rotated. The acquisition time for each scan is indicated on the horizontal axis.
• Figure 3: Left: Regular linear sampling pattern, as used for the static validation data. Right: CASPR sampling pattern, as used for the dynamic data and the hybrid dynamic/validation data. The readout direction ($k_x$, not shown) is perpendicular to the figure. The colors indicate the sampling order for the first 128 readouts.
• Figure 4: Overview of the different scans in the dataset. From top to bottom: high-resolution anatomical image; out-of-phase image from the DIXON scan with the muscle segmentation masks overlaid; fifth validation image from the static validation data (at the maximum pressure level of 80 mmHg); image at the maximum deformation reconstructed from the validation part of the hybrid dynamic/validation scan; time-averaged image from the data of all six dynamic repetitions of dynamic scan 1. Note that due to the radiological convention, the right leg is shown on the left side of the transverse and coronal images.
• Figure 5: Binned image reconstruction approach. Projections on the readout axis are calculated for each validation image and each CASPR shot of dynamic scan 1. The correlations between these projections are then used to divide the shots into nine bins, with each bin corresponding to one validation image. The first principal component of the projection data is used as motion surrogate signal. The validation images and corresponding images from the binning strategy are shown for bin 1 (no deformation) and bin 5 (maximal deformation).