Lea Behrendt1,2, Andreas Voskrebenzev1,2, Filip Klimeš1,2, Marcel Gutberlet1,2, Hinrich Winther1, Till Kaireit1,2, Tawfik Moher Alsady1,2, Gesa Pöhler1,2, Frank Wacker1,2, and Jens Vogel-Claussen1,2
1Institute of Diagnostic and Interventional Radiology, Hannover Medical School, Hannover, Germany, 2Biomedical Research in Endstage and Obstructive Lung Disease Hannover (BREATH), German Center for Lung Research (DZL), Hannover, Germany
Synopsis
Phase-resolved functional lung
(PREFUL) MRI is a promising contrast agent-free 1H method to derive information
about perfusion dynamics in free breathing. The analysis includes the
segmentation of a main vessel to determine the perfusion phase for
reconstruction of a full cardiac cycle. Since manual segmentation can be
challenging and time-consuming, an algorithm for automated phase sorting is
desirable. In this study such an algorithm is proposed, tested in 34 patients
with CF and CTEPH and compared to manual phase sorting. The algorithm was able
to perform segmentation in all cases and showed improved phase sorting compared
to manual segmentation.
Introduction
Phase
resolved functional lung (PREFUL) MRI1 is a contrast agent free 1H
technique, allowing for simultaneous assessment of lung ventilation2
and perfusion3 dynamics. It is based on the voxel-wise time series
analysis of dynamic lung MRI in free breathing4. To derive dynamic
information about pulmonary ventilation and perfusion, lung images are
co-registered, low-pass or high-pass filtered to extract the ventilation or
perfusion components and retrospectively sorted to reconstruct one respiratory
and cardiac cycle.
During PREFUL
postprocessing, the cardiac phase is determined using the mean signal time
series in a manual segmented region of interest (ROI) inside a main pulmonary
vessel1. However, manual segmentation is time-consuming, requires
user interaction and is challenging especially in posterior and anterior slices
of the lung. Therefore, in this study, an algorithm for automated cardiac phase
sorting is proposed. The applicability of the developed algorithm was tested in
patients with chronic lung disease in slices covering the whole lung.Method
14
patients with cystic fibrosis (CF) and 20 patients with chronic thromboembolic
pulmonary hypertension (CTEPH) were included in this study. Imaging was
performed on a 1.5T scanner (Avanto or Aera, Siemens Healthcare, Erlangen,
Germany) using a spoiled gradient echo sequence with the following settings:
Field of view = 225 x 225mm2 - 500 x 500mm2, matrix size
128 x 96 - 128 x 128 (all interpolated to 256 x 256), slice thickness 15mm,
slice spacing 15mm - 30mm, TE = 0.67ms - 0.91ms, TR = 3ms, flip angle 5° - 8°,
temporal resolution 192ms - 289ms. Between 4 and 13 coronal slices covering the
whole lung were obtained for each patient.
All
images were registered towards one fixed image in intermediate lung position5.
Afterwards, an automated algorithm for cardiac phase sorting inside PREFUL1
was implemented (Figure 1): Therefore, an ROI (Rsort), consisting of
at least one voxel cluster, located inside a pulmonary vessel or the heart
shall be segmented automatically. First, after automatic segmentation of the
lung parenchyma using a semantic convolutional neural network6, a
search ROI (As) consisting of both lungs and the mediastinum was
created. Perfused vessels and the heart, provide a high signal intensity and a
large amplitude in the cardiac frequency spectrum. Therefore, voxel clusters
inside vessels and the heart were identified from a temporal maximum intensity
projection map MMIP and a standard deviation map Mstd of
the high-pass filtered (cut-off frequency: 0.75Hz) dynamic image series using
the 98th percentile as
the lower threshold. To avoid cardiac phase shifts between the voxel clusters
and to improve the sine fit during reconstruction of the cardiac cycle, Rsort
was iteratively adjusted by comparing the goodness of fit parameter R2
of the acquisitions fitted to the reconstructed sine wave before and after
removing a cluster (Figure 2). If the sine fit improves, the cluster will be
removed from Rsort. As a lower limit, at least one cluster had to
remain in Rsort. Finally, images were sorted according to their
perfusion phase and interpolated to 15 phases at an equidistant time grid
covering one cardiac cycle.
R2 values
of the sine fit were calculated for all slices and compared to those obtained
with manually segmented ROIs inside a main vessel using a paired two-sided
Wilcoxon test.Results
For
every slice, a sorting ROI inside a vessel or the heart was found by the
algorithm. In Figure 3, segmented ROIs, corresponding sine fits and
perfusion-weighted maps obtained using the automatic algorithm and manual
segmentation are shown for selected slices of a CTEPH patient.
Perfusion-weighted maps obtained by the algorithm are comparable or, especially
for anterior slices without large vessels, improved to those obtained by manual
segmentation. R2 values for all slices are shown in Table 1.
Including all patients, median R2 increased from 0.81 (0.71-0.88)
using manual segmentation to 0.85 (0.79-0.90) using the automatic algorithm
(P<0.001). For the CF patients, R2 increased from 0.86
(0.75-0.91) to 0.88 (0.83-0.91) (P<0.001) and for the CTEPH patients, R2
increased from 0.79 (0.70-0.85) to 0.84 (0.77-0.89) (P<0.001).Discussion
Applicability
of automated cardiac phase sorting was successfully tested in CF and CTEPH
patients with similar or improved results of further PREFUL analysis compared
to manual segmentation. Even at posterior and anterior slices containing only
small vessels, where manual segmentation is difficult, the algorithm could
still segment a proper ROI inside a vessel or the heart. Limitations of the
algorithm occurred solely for peripheral posterior slices, where additional
spinal ROI clusters were found. However, for those slices, manual segmentation
was almost unfeasible. Further, functional maps generated for those slices, are
often not diagnostically conclusive due to through-plane motion and partial volume
effects.
Considering
the significantly increased coefficient of determination in comparison to the
manual evaluation, the algorithm may help to improve pulmonary perfusion
assessment using PREFUL.Conclusion
The
algorithm presented in this study allows for automated cardiac phase sorting of
coronal slices covering the whole lung during PREFUL postprocessing. This is an
important step towards a completely automated PREFUL analysis.Acknowledgements
No acknowledgement found.References
-
Voskrebenzev A, Gutberlet M, Klimeš F, et al. Feasibility of quantitative
regional ventilation and perfusion mapping with phase-resolved functional lung
(PREFUL) MRI in healthy volunteers and COPD, CTEPH, and CF patients. Magn Reson
Med 2018;79:2306-2314.
- Klimeš F,
Voskrebenzev A, Gutberlet M, et al. Free‐breathing quantification of regional ventilation derived by phase‐resolved functional lung (PREFUL) MRI. NMR in Biomedicine 2019;32:e4088.
- Kaireit TF, Voskrebenzev A, Gutberlet M, et al. Comparison of quantitative
regional perfusion-weighted phase resolved functional lung (PREFUL) MRI with
dynamic gadolinium-enhanced regional pulmonary perfusion MRI in COPD patients.
J Magn Reson Imaging 2019;49:1122-1132.
- Bauman
G, Puderbach M, Deimling M, et a. Non-contrast-enhanced perfusion and
ventilation assessment of the human lung by means of Fourier decomposition in
proton MRI. Magn Reson Med 2009;62:656-664
- Voskrebenzev A, Gutberlet M, Kaireit TF, et al. Low-pass imaging of dynamic
acquisitions (LIDA) with a group-oriented registration (GOREG) for proton MR
imaging of lung ventilation. Magn Reson Med 2017;78:1496-1505.
- Winther H, Gutberlet M, Hundt C, et al. Deep
semantic lung segmentation for tracking potential pulmonary perfusion
biomarkers in chronic obstructive pulmonary disease (COPD): The multi‐ethnic study of atherosclerosis COPD study. J Magn
Reson Imaging doi: 10.1002/jmri.26853