Amirah Faisal Alsaedi1,2, Jasmina Panovska-Griffiths3, Xavier Golay2, and Sotirios Bisdas2,4
1Department of Radiology Technology, Taibah University, Medina, Saudi Arabia, 2Department of Brain Repair & Rehabilitation, UCL, Queen Square, Institute of Neurology, London, United Kingdom, 3Department of Applied Health Research, UCL, London, United Kingdom, 4Lysholm Department of Neuroradiology, The National Hospital for Neurology and Neurosurgery, University College Hospitals NHS Trust, London, United Kingdom
Synopsis
This study aimed to assess the diagnostic
performance of multiparametric MRI radiomics for glioma class prediction
according to the WHO 2016 classification. Histogram features were extracted
from prospectively acquired multiparametric MRI (pCASL, DSC-MRI, DCE-MRI, and
DWI) in 32 patients with primary gliomas. The uncombined significant features
of ASL, ADC, DSC, and DCE, revealed diagnostic performances varying from low
(44% ) to fair (86%) and unable to predict all the histomolecular classes.
However, combining them for each MRI method, independently, enhanced the
diagnostic accuracy up to 100% and predict all the classes. This alludes the use
of multimodal radiomics for glioma classification.
Introduction
Currently,
histological biopsy is the gold standard for classifying gliomas according to the
most recent histomolecular features. However, this process is both invasive and
challenging when the lesion is at eloquent brain regions. Considering the interaction
between the presence of the IDH-mutation, the upregulation of the hypoxia
induced factor (HIF), angiogenesis and increased cellularity1, perfusion and diffusion MRI may be able to
predict the presence of such mutations indirectly. Recently, several studies
reported the subsidiary role of both perfusion and diffusion MRI2–5 in prediction of gliomas histomolecular class.
However, the extracted values are often evaluated as a single value and
mostly from the ‘hot-spot’ region of interest. In light of tumour
heterogeneity, considering the signal intensity distribution in the segmented entire
tumour volume via histogram analysis could enable more precise diagnosis free
of inter-observer variability6 7. Useful, but “hidden” from the human eye
information, can be extracted from the image using histogram texture analysis forming
the basis of radiomics. This creates high dimensional data that can subsequently
be reduced to ensure enhanced accuracy
and elimination of redundant elements. This study aims to assess the
diagnostic performance of multi-parametric MRI radiomics from the entire tumour
volume using subsequently feature reduction and combination of the standalone
modalities for classification of primary gliomas. Methods
Thirty-two adults with untreated gliomas were
prospectively recruited and underwent multimodal MRI perfusion with
pseudo-continuous arterial spin labelling (pCASL), dynamic susceptibility
contrast-enhanced (DSC), and dynamic contrast-enhanced (DCE) MRI, as well as diffusion-weighted
imaging (DWI). The gliomas were subsequently biopsied or surgically removed and
the final diagnosis was established using the WHO 2016 classification. DCE maps
were generated using both the modified Tofts model (mTK)8 and the Lawrence and Lee model (L&L)9. The entire tumour was segmented manually on
FLAIR, while grey matter (GM) was used as an internal reference, being automatically
segmented on a high-resolution T1-weighted volume. Histogram features (mean,
standard deviation, 95th-percentile, kurtosis, skewness, median, inter-quartile
range, mode, minimum, maximum, variance, entropy, median z-score, and slope of
the cumulative distribution function) were extracted from absolute and relative
(normalised) values (aT, rT, respectively). Additional percentiles were extracted
from the ADC map, 10tile to 90tile with 10 increment. Based on the Kruskal-Wallis
H test, only the statistically significant features were kept. These features were
further reduced using pairwise correlation (PC) and backward elimination (BE). The
final diagnostic performance was assessed using multinomial logistic regression,
both for individual features as well as for combined features from each modality.
The weighted average F1 (F) being used as indicator to reflect the
classification accuracy for the imbalanced sample sizes.Results
Figure 1 shows the diagnostic performance of
the significant features separately and combined. The uncombined significant
features of ASL, ADC, DSC, and DCE, revealed diagnostic performances varying
from low (44% ) to fair (86%) and unable to predict all the histomolecular
classes. However, linking the significant histogram features, mostly from rT,
of each MRI method, independently, enhanced the diagnostic accuracy up to 100% and
made it feasible to predict all the classes. As an individual technique, DSC
showed superior diagnostic performance (overall 100%), though utilising both PCASL
and ADC has similar accuracy (100%) to it. The DCE-L&L performed better
than ASL (F range: 87% to 100%, 78% to 93%, respectively), but not the DCE-mTK
(F range: 64% to 83). Notably, DCE-L&L attained identical diagnostic
performance as DSC in term of the molecular classification. Discussion
Both ADC and DSC are widely used in current
clinical glioma MRI protocols10. Based on our results, DSC might be omitted
when using the combined histogram features from ADC while still achieving
effective diagnoses. However, ADC does not provide information regarding the
tumour’s hemodynamic nature.
In contrast, though ASL has still not been
fully adopted in routine glioma MRI scanning, our results revealed that the
combined histogram features from PCASL and ADC had comparable diagnostic
performance to DSC (overall 100%), confirming that DSC could be here safely
skipped, something that would be important in light of the increasing awareness
for gadolinium deposition, with the caveat of longer scanning time required for
ASL and limited dissemination of PCASL. Nonetheless, out of the PCASL with
multiple TIs acquired in 5min 9s in our study, only the ASL data at the 6th
TI was considered for post-processing and analysis, rendering the actual PCASL
scanning time about 1min 2s, similar to the DSC acquisition time (1min 26s).
A critical step in radiomics for diagnostic
purposes is feature reduction11. The reduction was performed here using PC and
BE11,12, where the latter demonstrated better results.
This might be because the PC method considers the relationship between both
feature-class and feature-features, while the BE considers all the significant
class-related features while ignoring feature-feature mutual information. These
results, however, could be excessively optimistic, due to the small sample size,
and larger cohorts are needed.Conclusion
Our preliminary investigation suggests that
either DSC as an individual modality or alternatively combined ASL and ADC have
similar and excellent accuracy as non-invasive biomarkers for gliomas classes
prediction. This reinforces the evidence for the diagnostic accuracy of DSC and
at the same time alludes to the use of multimodal radiomics for glioma
classification. Acknowledgements
No acknowledgement found.References
1. Cairns RA, Harris I, Mccracken S, Mak TW.
Cancer cell metabolism. Cold Spring Harb Symp Quant Biol.
2011;76(December 2011):299-311. doi:10.1101/sqb.2011.76.012856.
2. Brendle C, Hempel J-M, Schittenhelm J, et
al. Glioma Grading and Determination of IDH Mutation Status and ATRX loss by
DCE and ASL Perfusion. Clin
Neuroradiol. 2017. doi:10.1007/s00062-017-0590-z.
3. Lin Y, Xing Z, She D, et al. IDH mutant
and 1p/19q co-deleted oligodendrogliomas: tumor grade stratification using
diffusion-, susceptibility-, and perfusion-weighted MRI. Neuroradiology.
2017;59(6):555-562. doi:10.1007/s00234-017-1839-6.
4. Xing XZ, Yang XX, She XD, Lin XY, Zhang
XY, Cao XD. Noninvasive Assessment of IDH Mutational Status in World Health
Organization Grade II and III Astrocytomas Using DWI and DSC-PWI Combined with
Conventional MR Imaging. 2017.
5. Latysheva A, Emblem KE, Brandal P, et al.
Dynamic susceptibility contrast and diffusion MR imaging identify
oligodendroglioma as defined by the 2016 WHO classification for brain tumors:
histogram analysis approach. Neuroradiology. 2019;61(5):545-555.
doi:10.1007/s00234-019-02173-5.
6. Just N. Improving tumour heterogeneity
MRI assessment with histograms. Br J Cancer. 2014;111(12):2205-2213.
doi:10.1038/bjc.2014.512.
7. Zeng Q, Jiang B, Shi F, Ling C, Dong F,
Zhang J. 3D pseudocontinuous arterial spin-labeling MR imaging in the
preoperative evaluation of gliomas. Am J Neuroradiol. 2017;38(10):1876-1883.
doi:10.3174/ajnr.A5299.
8. Tofts PS. I Modeling Tracer Kinetics in
Dynamic. J Magn Reson Imaging. 1997;7(1):91-101.
9. Lawrence KS St., Lee T-Y. An Adiabatic
Approximation to the Tissue Homogeneity Model for Water Exchange in the Brain:
II. Experimental Validation. J Cereb Blood Flow Metab.
1998;18(12):1378-1385. doi:10.1097/00004647-199812000-00012.
10. Thust SC, Heiland S, Falini A, et al.
Glioma imaging in Europe: A survey of 220 centres and recommendations for best
clinical practice. Eur Radiol. 2018;28(8):3306-3317.
doi:10.1007/s00330-018-5314-5.
11. Afshar P, Mohammadi A, Plataniotis KN,
Oikonomou A, Benali H. From handcrafted to deep-learning-based cancer
radiomics: Challenges and opportunities. IEEE Signal Process Mag.
2019;36(4):132-160. doi:10.1109/MSP.2019.2900993.
12. Hall MA, Smith LA. Correlation-based Filter
Approach vs. Wrapper. 1995.