Kerstin Lagerstrand1,2, Hanna Hebelka1,3, Leif Thorén1,3, Christian Waldenberg1,2, and Helena Brisby1,4
1Institute of Clinical Sciences, Gothenburg University, Gothenburg, Sweden, 2Medical Physics and Biomedical Engineering, Sahlgrenska University Hospital, Gothenburg, Sweden, 3Radiology, Sahlgrenska University Hospital, Gothenburg, Sweden, 4Orthopaedics, Sahlgrenska University Hospital, Gothenburg, Sweden
Synopsis
Imaging-based features are needed to improve the characterization of
degenerative IVD-changes and possibility of
finding a linkage between features and pain.
Multiple T2w-imaging-features and Machine-Learning was used for
classification of fissures involving outer annulus and for pain-positive discograms.
Fissures
were classified with high accuracy/precision using regional/heterogeneity
features with/without axial loading of the spine.
For
pain-positive discograms, a larger number of such MRI-features contributed to
the classification.
Findings suggest that multiple MRI-features, extracted from T2w-imaging,
improve the classifications, and that regional/heterogeneity features extracted
with both conventional imaging with the spine
unloaded and with axial loading of the spine are of importance.
Introduction
To date, several studies have shown
the value of quantitative MRI and axial loading during MRI (alMRI) for
detection of degeneration-related IVD-changes using single parametric features,
e.g. mean IVD T2-value measured with conventional MRI in the supine position
(uMRI) or the separation between histogram peaks representing nucleus pulposus
and annular fibrosis. However, these imaging features neither correlate with
clinical symptoms, nor do they consistently predict the development of low back
pain (LBP).1,2
In this study, we examine the value of multiple MRI-features using Machine
Learning techniques for improved characterization of degenerative IVD-changes,
but also for the possibility of finding a linkage between MRI-features and
pain.Methods
During one single day, MRI, CT and low-pressure
discography (<50psi) were performed on the lumbar spine of each 30
LBP-patients. Quantitative MRI-features, which previously have been shown to
correlate with IVD degeneration, were extracted from sagittal T2-weighted (T2w)
images, acquired on a 1.5T system (Siemens Magnetom Symphony Maestro Class,
Erlangen, Germany) using both uMRI and alMRI (TR=4ms,TE=124ms,slice
thickness=4mm). Axial load was created with a compression device (DynaWell,
Dynawell diagnostics AB, Las Vegas, NV USA) configured to generate axial
compression of the lumbar spine, corresponding to half the patient's body
weight.
For each IVD, global and regional MRI-features,
as well as heterogeneity features were extracted (Table 1) using an in-house
segmentation program based on MatLab (R2018b, Mathworks, Natick, Massachusetts,
U.S.A.). Global features were extracted from the entire IVD using regions of
interest (ROIs) that were drawn semi-automatically on the 5 central slices of
the T2w-images. Regional features were extracted from volumetric IVDs, dividing
the ROIs into 5 equally large parts in the sagittal direction (ROI1 anterior to
ROI5 posterior). The standard deviation of the mean and the difference between
histogram peaks representing the nucleus pulposus and annular fibrosis were
used as measures of IVD-heterogeneity.
Discograms were performed in 84 IVDs and labeled
as pain-positive for concordant pain response at a pressure <50psi. Also,
all CT-discograms were labeled by the reader according to the Dallas Discogram
Description (DDD) and Adams grading, and then digitomized into annular fissures
involving the outer annulus or not, i.e. DDD≥2/DDD≤1 respectively
Adams≥D/Adams≤C.
The multiparametric analysis was performed using
Anaconda with the Python distribution (Anaconda Software Distribution. Computer
software. Vers. 2-2.4.0. Anaconda, Nov. 2016). Supervised “Decision Tree” Machine
Learning was used to examine how well the included MRI-features classified the
“true labels” DDD≥2/DDD≤1 and Adams≥D/Adams≤C, as well as pain-positive
discograms. A 10-fold cross validation was used for model derivation and the
supervised classifiers was applied to the remaining third of data in a separate
data file containing unseen data. All data was normalized for quantification.
Results are presented as Decision Trees,
confusion matrices and classification reports in terms of confusion matrix, accuracy,
precision, recall, and F1 score (harmonic mean of the precision and recall).Results
Figure 1 shows the evaluation of the
“Decision Tree” algorithm for classification of fissures involving outer
annulus (same results for DDD and Adams) and pain-positive discograms.
Corresponding “Decision Trees” are shown in Figure 2. As can be seen in the
figures, the algorithm classified IVDs with fissures involving the outer
annulus with high accuracy and precision using only a few regional
MRI-features, while a large number of regional and heterogeneity features were
needed for the classification of pain-positive discograms, still, resulting in
a lower accuracy and precision.
Results from the cross validations showed high
reproducibility for the fissure classification, mean:0.9 (std:0.1) and lower
reproducibility for pain-positive discograms, mean:0.7 (std:0.1).Discussion
This study suggests that multiple
MRI-features, extracted from T2w-images, improve the classification of annular
fissures and pain-positive discograms, and that regional as
well as heterogeneity features using both uMRI and alMRI are of importance for
the classifications.
Especially, MRI-features that reflect
IVD-behaviors during alMRI and alMRI-uMRI at the position of the nucleus
pulposus (ROI3) and the transition zone between the nuclear pulposus and
annular fibrosis (ROI4) were shown to be important for the fissure
classification. Moreover, the heterogeneity of the IVD during uMRI at the
position of the nucleus pulposus added a small contribution to the
classification. The importance of adding image information from the central
regions of the IVD for fissure classification may be explained by the
associated remodulation of the nucleus pulposus.3,4
Regional and heterogeneity information seemed to
be of importance also for the classification of pain-positive
discograms. Especially, the heterogeneity of
the IVD during alMRI at ROI4 and the load induced IVD-change at the most
anterior and posterior sub-volume, i.e. ROI1 and ROI5, contributed the most to
the pain classification. Thus, pain-positive discograms
seem to be characterized by a significant loading effect on the annular
fibrosis, but more importantly, by the variation in the IVD matrix structure at
the transition zone, enhanced by alMRI.Conclusion
This study demonstrates that
multiparametric MRI improves the characterization of IVD fissures involving the
outer annulus and pain-positive discograms and, thus, have an important value in future
LBP-research. To enable direct clinical usability, additional features linked
to pain needs to be added in order to improve detection of pain-positive spinal
segments.Acknowledgements
No acknowledgement found.References
1. Doniselli et al. Eur Spine J 2018;27(11):2781-2790
2. Vagaska et al. Medicine (Baltimore) 2019;98(17):e15377.
3. Waldenberg et al. PLOS ONE 14(8)
4. Sharma et al. Spine (Phila Pa 1976). 2011;36(21):1794–800