Rapid Fire Abstracts
Daniel Amsel, MSc
PhD Student
University of Tuebingen, Tuebingen, Germany, Germany
Daniel Amsel, MSc
PhD Student
University of Tuebingen, Tuebingen, Germany, Germany
Jens Wetzl, PhD
Applications Developer
Siemens Healthineers, Germany
Rolf Gebker, MD, PhD
Cardiologist
German Heart Center Berlin, Germany
Christoph Tillmanns
Cardiologist, Head of CardioImaging
Diagnostikum Berlin, Germany
Kelvin Chow, PhD
MR Collaboration Scientist
Siemens Healthcare Ltd., Canada, Canada
Daniel Giese, PhD
Scientist
Siemens Healthineers, Germany
Thomas Küstner, PhD
Professor
University Hospital of Tübingen
In recent years cardiac T1 mapping has gained popularity in both research and clinical practice and is used to diagnose conditions like diffuse fibrosis und cardiomyopathies1. Nevertheless, the commonly used MOLLI2 acquisition scheme limits the spatial resolution of resulting T1 maps, thereby impeding the examination of small lesions. In this work, a deep learning-based image reconstruction technique is utilized to allow for the acquisition of T1 maps with higher resolution.
Methods:
An end-to-end variational network3 (VarNet) utilizing 2D+t convolutions and patchwise squeeze and excitation layers4 was trained to reconstruct undersampled MOLLI inversion recovery images. Using the FIRE5 framework, a prototype reconstruction was integrated inline into an existing vendor software for T1 mapping.
A dataset consisting of 1750 MOLLI 5(3)3 acquisitions with an acquired resolution of 1.4x2.1 mm2 (interpolated to 1.4x1.4 mm2) was used for training, validation, and retrospective testing. The data was acquired using a vendor T1 mapping sequence with an acceleration rate of R=2. The corresponding GRAPPA6-interpolated k-space was regarded as ground truth and retrospectively undersampled using a k-t-sampling pattern7 with R=4.
Additionally, 10 MOLLI 5(3)3 scans were acquired prospectively in healthy subjects using the above-mentioned sequence as well as a research sequence with R=4 and an increased resolution of 1.1x1.1 mm2 (acquired and reconstructed).
Training and prospective test data were acquired on clinical 3T scanners (MAGNETOM Skyra, Vida, Lumina, Siemens Healthineers AG, Forchheim, Germany) at two different sites.
Results:
The metric scores for comparing VarNet reconstructions to ground truth inversion recovery images from the test dataset are displayed in Figure 1a). The scores indicate good reconstruction performance.
This is also reflected in the Bland-Altman plot (Figure 1b) illustrating the mean T1 difference in the myocardium. A low mean difference of 6.7 ms and narrow limits of agreement (LoA) of 51.8 ms and -38.4 ms indicate good T1 agreement. Good performance is maintained when changing to prospectively acquired data. Mean T1 differences across the six mid-ventricular AHA segments in the prospectively acquired data (Figure 1c) show a high degree of agreement with values reported for the vendor reconstruction (Mean: -10.9ms; LoA: 59.2 ms and -81.2 ms). Remaining stronger deviations are likely due to imperfect segmentation and partial volume effects in GRAPPA-reconstructed T1 maps.
In a patient diagnosed with myocardial infarction, the T1 maps using GRAPPA and the VarNet for reconstruction both clearly depict the corresponding lesion (Figure 2). No meaningful T1 differences are observed.
Exemplary T1 maps for prospectively acquired data (Figure 3) show improved image sharpness and visibility of small structures when using the VarNet compared to GRAPPA for reconstruction.
Conclusion:
A modified VarNet was proposed, that when integrated into an existing T1 mapping reconstruction, allows for the acquisition of higher resolution T1 maps without increasing the scan or breath hold duration. Future clinical evaluation will investigate this method’s impact on the detection of small lesions.
