2133
Volume-independent radiomic features from T2w-FLAIR MRI could reveal mutation of histones in diffuse intrinsic pontine glioma1IMIV, Inserm, CEA, CNRS, Université Paris-Sud, Université Paris-Saclay, Orsay, France, 2Pediatric Radiology, Hôpital Necker Enfants Malades, AP-HP, Paris, France, 3UNATI, Neurospin, CEA, Université Paris-Saclay, Gif-sur-Yvette, France, 4Pediatric Neurosurgery, Hôpital Necker Enfants Malades, AP-HP, Paris, France, 5Cancérologie de l'enfant et de l'adolescent, Gustave Roussy, CNRS UMR 8203, Université Paris-Saclay, Villejuif, France
Synopsis
In diffuse intrinsic pontine glioma, the mutations of histones (H3.1 versus H3.3) are correlated with patient survival. A new method to compute radiomic features free of tumor volume effect was applied to four structural MR modalities and patients were classified according to histone mutation. The tumor was scanned by a 5 mm radius sphere and textural indices were computed inside each position. A total of 37 features calculated from T2w-FLAIR yielded an area under the Receiver Operating Characteristics curve greater than 0.85. T2w-FLAIR appears to be the most informative modality to predict mutation type.
Introduction
Diffuse intrinsic pontine glioma (DIPG) is a rare inoperable lethal pediatric cancer1, mainly diagnosed with magnetic resonance imaging (MRI). Recent studies showed that the two main DIPG histone mutations (H3.1 and H3.3) were correlated with patient overall survival and that the median age at diagnosis was significantly smaller in H3.1 in comparison with H3.3 mutated patients2. Radiomics relies on quantitative image features computation for prognosis or prediction tasks and has been increasingly used in tumor analysis3–6. However the actual informative contribution of those numerous features is still discussed, and textural indices can be strongly biased by the volume of the region of interest (ROI) in which they are computed7. The present work proposes a method to compute radiomic features free of tumor volume effect. These features were computed in four different structural MRI modalities and their ability to classify the genetic mutation of DIPG (H3.1 or H3.3) was studied.Materials and Methods
A retrospective analysis of 28 DIPG patients (8 H3.1/20 H3.3, mean age 8±3 years) was performed. Structural diagnosis MR scans: T1-weighted (T1w), T2-weighted (T2w), T1-weighted post-contrast injection (T1w-c), T2-weighted FLAIR (T2w-FLAIR) were included. The complete processing workflow is presented in figure 1. Images were processed using an intensity standardization pipeline8 and resampled to isotropic voxels (1x1x1 mm3). For each patient, a large spherical ROI (2.3 to 20.5 cm3) was drawn inside the tumor (sT). A small 5 mm radius sphere (s5) scanned the sT, with 1 voxel step, and 79 textural indices were computed for each s5 position using PyRadiomics9. A total of 19 indices were computed from first-order histograms, and using a fixed intensity bin size (2), 28 indices were computed from the gray-level co-occurrence matrix (GLCM), 16 from the gray-level run length matrix (GLRLM), and 16 from the gray-level size zone matrix (GLSZM). The mean of each index inside sT were computed leading to 316 (79 textural indices x 4 modalities) imaging features characterizing the DIPG. The same feature computation procedure was employed in two additional scenarios: using sT (a) dilated and (b) eroded with a 5-mm radius sphere.Statistical analysis
The Spearman’s correlation coefficient (r) between each image feature and the ROI volume was estimated. The area under the Receiver Operating Characteristics curve (ROC–AUC) was used to evaluate the H3.1/H3.3 discriminative power of age at diagnosis and of each imaging feature. Using the three definitions of tumor ROI (initial, eroded and dilated) robustness of imaging features to the ROI segmentation was evaluated with the absolute agreement intraclass correlation10 (ICC).Results
3D maps of textural indices within one patient’s tumor are presented in figure 2. None of the 316 imaging features was highly correlated with volume (greatest |r| = 0.41 for the mean of T1w-c first-order 10percentile). A total of 37 imaging features derived from T2w-FLAIR, 2 from T1w-c, and none from T2w and T1w presented an ROC–AUC greater than or equal to 0.85 (figure 3). Six features (figure 3) showed a better score than age at diagnosis (ROC-AUC = 0.888). The boxplots of age at diagnosis in addition to five H3.1/H3.3 discriminants imaging features are shown in figure 4. All T2w-FLAIR and T1w-c features, (except T1w-c first-order Kurtosis, T2w-FLAIR first-order Kurtosis T2w-FLAIR GLCM ClusterShade, T2w-FLAIR first-order Skewness) had an ICC greater than 0.75.Discussion
The imaging features presented in this work are free of volume bias as textural indices are computed within ROI of constant volume (s5). Our results show that some of these features are related to the histopathological results on the DIPG genetic mutation H3.1/H3.3, presenting ROC-AUC equivalent to (or greater than) age at diagnosis in our series. Notably, quantitative imaging features from T2w-FLAIR are the most discriminant, while no features computed from T2w and T1w resulted in great ROC-AUC, which is consistent with previous results11. Furthermore, all discriminant features computed with the proposed method are robust to changes in ROI delineation according to the estimated ICC values.Conclusion
Quantitative volume-independent radiomic features describing imaging heterogeneities provide information regarding the genetic mutation of DIPG when calculated from T2w-FLAIR scans.Acknowledgements
No acknowledgement found.References
1. Cohen, K. J., Jabado, N. & Grill, J. Diffuse intrinsic pontine gliomas—current management and new biologic insights. Is there a glimmer of hope? Neuro. Oncol. 19, 1025–1034 (2017).
2. Castel, D. et al. Histone H3F3A and HIST1H3B K27M mutations define two subgroups of diffuse intrinsic pontine gliomas with different prognosis and phenotypes. Acta Neuropathol. 130, 815–827 (2015).
3. Lambin, P. et al. Radiomics: Extracting more information from medical images using advanced feature analysis. Eur. J. Cancer 48, 441–446 (2012).
4. Kickingereder, P. et al. Radiogenomics of Glioblastoma : Machine Learning – based Classification of Molecular Characteristics by Using Multiparametric and Mutiregional MR Imaging Features. Radiology 281, 907–918 (2016).
5. Aerts, H. J. W. L. The Potential of Radiomic-Based Phenotyping in Precision Medicine. JAMA Oncol. 2, 1636–1642 (2016).
6. Gillies, R. J., Kinahan, P. E. & Hricak, H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 278, 563–577 (2016).
7. Goya Outi J. et al. Impact of volume size on radiomic features computed from MR images: an illustration in pediatric neuroimaging. MAGMA 30, S648 (2017).
8. Goya Outi J. et al. A pipeline to standardize intensities in MR images for reliable longitudinal studies in neuro-oncology. MAGMA 30, S335 (2017).
9. van Griethuysen J.J.M. et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 77(21), e104–e107(2017).
10. McGraw K.O., Wong S. Forming inferences about some interclass correlation coefficients. Psycol Methods 1(1):30-46 (1996). 11. Hargrave D., Chuang N., Conventional MRI cannot predict survival in childhood diffuse intrinsic pontine glioma. J. Neurooncol 86, 313-319 (2008).
11. Hargrave D., Chuang N., Conventional MRI cannot predict survival in childhood diffuse intrinsic pontine glioma. J. Neurooncol 86, 313-319 (2008).
Figures
