Grouping gliomas using the telomerase reverse transcriptase (TERT) gene and IDH mutations, and 1p/19q co-deletion status was demonstrated to be useful previously for clinical decisions. MR based radiogenomics might potentially be advantageous.
The aim of this study was to determine for the first time whether full distributions of the fractional anisotropy, relative anisotropy and ADC in normal appearing white matter were adequate predictors for machine learning algorithms to classify molecular subgroups based on TERT, IDH and 1p/19q co-deletion information.
Grouping gliomas using the telomerase reverse transcriptase (TERT) gene and IDH mutations, and 1p/19q co-deletion status was demonstrated to be useful previously for clinical decisions on gliomas1 which resulted in the inclusion of isocitrate dehydrogenase (IDH) mutation status within the 2016 WHO classification2.
Non-invasively obtaining molecular data using magnetic resonance (MR) based radiogenomics3 might potentially be advantageous. Diffusion anisotropy4 provides relevant information on tissue microstructure as a cancer biomarker. Some features (e.g. peak height, peak location, percentile values etc.) of the distribution of apparent diffusion coefficient (ADC) within the tumor were investigated for predicting low-grade glioma histological subtypes5. Likewise, features of ADC distribution in the normal appearing white matter (NAWM) of glioma patients were used for comparing non-infiltrative (meningiomas) versus infiltrative disease (gliomas).
The aim of this study was to determine for the first time whether full distributions of the fractional anisotropy (FA), relative anisotropy (RA) and ADC4 in NAWM were adequate predictors for machine learning algorithms (MLA) to classify molecular subgroups based on TERT, IDH and 1p/19q co-deletion information.
In this IRB approved study, cohorts were chosen out of consecutive 170 glioblastoma patients with written consent. Only monolateral patients with diffusion tensor MRI (TE/TR=89.1/7733ms, 60 slices, b=800s/mm2, 20 directions,) and T2W images (turbo spin echo, TE/TR=107/3470ms, 20 slices) collected on a 3T Siemens TrioTrim were selected. Molecular subtype data were obtained using Sanger sequencing with an Applied Biosystems™ 3500 Series Genetic Analyzer.
A cohort of 34 (17F/17M) patients of 41.82±13.87 years had full molecular subtype and imaging data. For this cohort, (TERT, IDH, co-deletion) status determined 6 groups: (---) triple negative, (+--) TERT only, (-+-) IDH only, (++-) TERT and IDH, [(--+), (+-+), (-++)] co-deletion enhanced, (+++) triple positive which had respectively 4, 5, 11, 2, 3, 9 patients.
A cohort of 59 (26F/33M) patients of 43.39±14.73 years had (TERT, IDH) and imaging data. For this cohort, 4 groupings were formed according to (TERT, IDH) status which contained 10 (--), 14 (+-) (TERT only), 24 (-+) (IDH only) and 11 (++) patients. From this cohort, (TERT only) and (IDH only) groups were also tested.
NAWM masks were delineated semi-automatically on T2W images using MIPAV6 software’s level set VOI routine. FSL©7 FLIRT© package calculated the transformation for co-registering B0 images onto T2W images subsequently applied to eigenvalue maps. In-house developed MATLAB™ routines were used for separating the hemispheres on the B0-NAWM masks and for calculating indices. The distributions in the NAWM and, both hemispheres were calculated and were normalized to have a unit integral and ranges [0 1] (FA, RA) and [0 3895 mm2/s] (ADC).
Histograms were obtained for 34-patients 20, 40, 60 and, for 59-patients 30, 60 and 90 bins. Each patient’s distribution was treated as a vector with number of bins elements. Total NAWM distributions were the first feature set and distributions in the contralateral side were subtracted from the healthy side to obtain the second feature set.
Mean centered feature vectors were predictors along with the molecular subgroups as labels for 22 MLAs in MATLAB™. MLA’s were run with leave-one-out and 50-fold cross-validation for 34-patient and 59 patient cohorts respectively, once without Principle Component Analysis (PCA) and also with PCA retaining 98% of the variance and 3 components were tested.
MLA’s for (TERT, IDH, co-deletion) classification resulted poorly with the top result from total NAWM RA, 20 bin, using linear support vector machines (SVM) without PCA with 47.1% accuracy.
For (TERT, IDH) top result was obtained with Coarse Tree, 30 bin, total NAWM RA, 3 component PCA resulting in 49.2% accuracy.
Top (TERT only) result was obtained with Ensemble Bagged Trees, 30 bin, total NAWM ADC, 3 component PCA resulting in 72.9% accuracy.
Top (IDH only) result was obtained with Coarse Tree, 30 bin, total NAWM ADC, without PCA resulting in 83.1% accuracy.
All of the classifications obtained from distribution differences resulted in closer but always poorer outcomes.
1. Eckel-Passow, J.E. et al. Glioma Groups Based on 1p/19q, IDH, and TERT Promoter Mutations in Tumors. New England Journal of Medicine 372, 2499-2508 (2015).
2. Louis, D.N. et al. The 2016 World Health Organization Classification of Tumors of the Central Nervous System: a summary. Acta Neuropathologica 131, 803-820 (2016).
3. Moton, S., Elbanan, M., Zinn, P.O. & Colen, R.R. Imaging Genomics of Glioblastoma: Biology, Biomarkers, and Breakthroughs. Topics in Magnetic Resonance Imaging 24, 155-163 (2015).
4. Basser, P.J. & Pierpaoli, C. Microstructural and Physiological Features of Tissues Elucidated by Quantitative-Diffusion-Tensor MRI. Journal of Magnetic Resonance, Series B 111, 209-219 (1996).
5. Tozer, D.J. et al. Apparent diffusion coefficient histograms may predict low-grade glioma subtype. NMR in Biomedicine 20, 49-57 (2007).
6. McAuliffe, M., Edn. 8.0.2 (2018-02-13) Medical Image Processing, Analysis and Visualization (National Institutes of Health, https://mipav.cit.nih.gov/index.php; 2018).
7. Jenkinson, M., Beckmann, C.F., Behrens, T.E.J., Woolrich, M.W. & Smith, S.M. FSL. NeuroImage 62, 782-790 (2012).