MRI textural features can differentiate pediatric ... · Niha Beig1, Ramon Correa1, Rajat Thawani1,...
Transcript of MRI textural features can differentiate pediatric ... · Niha Beig1, Ramon Correa1, Rajat Thawani1,...
Dataset Description:
A retrospective cohort consisting of 59 MRI studies (Male = 30, Female = 29 ) was acquired from The University
Hospitals between 2003 – 2016.
The tumors were pathologically classified (MB = 22, EP = 12, Gliomas = 25) based on 2016 WHO Classification
of Tumors of CNS. [10]
The training set consisted of 30 patients (MB = 11, EP = 6, Gliomas = 13) and the independent validation set
comprised of 29 patients (MB = 11, EP = 6, Gliomas = 12)
• Wilcoxon rank sum test implemented to determine the top 10 discriminable features within training set. (p < 0.05)
• Random forest (RF) classifier model with a randomized 3-fold cross-validation was developed on this dataset.
• Using this model, an independent test set was validated via area under receiver operating characteristic (AUC).
52 radiomic features
capturing lesion
heterogeneity (CoLlAGe)
were extracted from the
entire region of interest
3Tesla MRI sequences :
• Gd-T1w,
• T2w,
• FLAIR
We presented a novel radiomics descriptor – CoLlAGe, based approach to identify non-invasive surrogate
markers of Posterior Fossa tumor types on routine T2w - MRI protocol.
CoLIAGe captures disorder in local arrangement of pixel-wise gradient orientations on radiologic imaging.
Medulloblastoma (aggressive brain tumor) reported higher CoLlAGe entropy values than Ependymomas and
Gliomas on MRI for pediatric brain tumor cases.
Our presented radiomic approach may allow for a noninvasive understanding of the various posterior fossa
tumor types as manifested on T2w images.
Future directions:
To validate our analysis on a multi-institutional independent dataset.
Background:
Medulloblastoma (MB), Ependymomas (EP), and Gliomas are common posterior fossa (PF) tumors, which
account for ~47% of all pediatric brain tumors. [1]
PF tumor types have different treatment regimens. MB are the most aggressive and require gross total
resection (GTR), radiation and chemotherapy whereas Gliomas are just monitored after GTR. [2]
Currently, histological findings are used to confirm the type of PF tumor via stereotactic biopsy. [3]
There is a clinical need to identify novel structural, and functional imaging based markers that can predict PF
tumor type in a non invasive preoperative setting – that can help with prognosis and treatment planning.
Previous Work:
Radiomics: The conversion of quantitative medical images into high-dimensional data that can be mined to
uncover the underlying pathophysiology. [4]
Radiomic features like Haralick capture 1st and 2nd order statistical patterns, while steerable filters like Gabor
capture dominant orientations that solicit a textural response.
Previous studies have used 2D and 3D based radiomic features to distinguish the PF types. But did not have
an independent validation set. [5,6]
Recently, we developed a new computerized descriptor, CoLlAGe, that distinguishes subtly different
pathologies in adult brain tumors. [7]
CoLlAGe captures co-occurrence of localized dominant orientations at pixel level, and brings together classical
Haralick and steerable filters by capturing statistics and frequency of local gradient orientations.
[1] Zakrzewski et al., “Posterior fossa tumours in children and adolescents. A clinicopathological study of 216 cases”, Folia Neuropathol 2003.
[2] Massimino et al., “Long-term results of combined pre-radiation chemotherapy and age-tailored radiotherapy doses for childhood medulloblastoma”, J
Neurooncol 2012.
[3] Chen et al., “Stereotactic biopsy for brainstem lesion: Comparison of approaches and reports of 10 cases”, J Chinese Medical Association 2011
[4] Gillies et al., “Radiomics: Images Are More than Pictures, They Are Data”, Radiology 2016.
[5] Orphanidou-Vlachoua et al., “Texture analysis of MR images and use of probabilistic neural network to discriminate posterior fossa tumors in children”,
NMR in Biomedicine 2014.
[6] Fetit et al., “3-D textural features of conventional MRI improve diagnostic classification of childhood brain tumors”, NMR in Biomedicine 2015.
[7] Prasanna et al., “Co-occurrence of Local Anisotropic Gradient Orientations (CoLlAGe): A new radiomics descriptor”, Sci Rep. 2016.
[8] Avants et al., “Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain.
Med Image Anal. 2008.
[9] Richards, J.E et al., A database of age-appropriate average MRI templates. Neuroimage, 2015
[10] Louis et al., “The 2016 World Health Organization Classification of Tumors of the Central Nervous System: a summary”, Acta Neuropathol 2016
Objective: Identify a set of radiomics (computer extracted) features on routine MRI sequences that can
distinguish the most common posterior fossa tumors - Medulloblastoma, Ependymomas, & Gliomas in children.
INTRODUCTION RESULTS
MATERIALS and METHODS
CONCLUSIONSREFERENCES
MRI textural features can differentiate pediatric posterior fossa tumors
Niha Beig1, Ramon Correa1, Rajat Thawani1, Prateek Prasanna1, Chaitra Badve2, Deborah Gold2, Anant Madabhushi1, Peter deBlank3, Pallavi Tiwari1
(1) Case Western Reserve University, Cleveland, OH(2) University Hospitals, Cleveland, OH
(3) Cincinnati Children's Hospital Medical Center, Cincinnati, OH
Lesion annotated by an
expert into four regions:
• Tumor necrosis
• Enhancing tumor
• Non enhancing tumor
• Edema (not shown)
T2w, and FLAIR were co-
registered to Gd-T1w MRI
using BSpline registration
and normalized using
pediatric atlases by age
[8,9]
Data acquisition and Feature extraction
Feature selection and classification
Hypothesis: We hypothesize that radiomic descriptor – CoLlAGe, can identify subtle
microarchitectural differences between MB, EP and Gliomas on routine pediatric MRI scans.
p = 5.75 x 10-4
Medulloblastoma Ependymoma
FIGURE 1 : Boxplot of CoLIAGe Sum Variance
FIGURE 2 : Scatterplot of the independent validation set in the feature space of CoLIAGe
Entropy and Sum Variance
FIGURE 3 : A – CoLIAGe sum variance for GliomaB – CoLlAGe sum variance for Ependymoma
C – CoLlAGe sum variance for Medulloblastoma
C
A
Result and Take-aways:
On the independent set, we found that sum variance and entropy of CoLlAGe (p< 0.002) on T2w protocol could
reliably distinguish MB from ( EP + Gliomas ) with an AUC of 0.88 (Accuracy = 0.82, Sensitivity = 0.63 and
Specificity = 0.94)
Sum variance of CoLlAGe possibly accounts for greater variation of scattered atypia and local accumulation of
mitotic processes as observed on histopathology. [7]
Higher entropy of CoLlAGe is indicative of more chaotic arrangement in areas of high viable cell population. [7]
Our results suggest that entropy of CoLlAGe from the entire tumor potentially captures high cellularity of MB
which enables its discrimination from ( EP + Gliomas ).
B
*
Glioma