ABSTRACT and FLAIR, and within each compartment:

ABSTRACT

BACKGROUND AND
PURPOSE: Activation of the AKT pathway has a significant role in cellular
proliferation, migration, and apoptosis, and is associated with poor prognosis
in GBM. Biopsy and genetic sequencing are used to diagnose this mutation, which
is not a routine procedure in the majority of hospitals and clinics. In this
work, we explored the feasibility of directional gradients (Gabor) and local
intensity statistics (Haralick, Law) texture features obtained from different
tumor-specific sub-compartments (enhancing, non-enhancing, infiltrating edges,
and necrotic regions) on T1, T2 and FLAIR sequence in non-invasively predicting
AKT pathway mutational status from routinely acquired MRI.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

 MATERIALS AND METHODS: A retrospective study of brain tumor MR
imaging performed on images from TCIA. 75 patient studies were analyzed. Brain
lesions on MR imaging were manually annotated by an expert neuroradiologist. A
set of three dimensional radiomic features was extracted for every lesion on
each MR imaging sequence: gadolinium T1WI, T2WI, and FLAIR, and within each
compartment: enhancing tumor, edema, necrosis. Rank-sum tests were used to find
features that correlated with the AKT pathway status of the patient studies.

RESULTS: 5
radiomic texture features with the most mutual information were Gabor, Sobel,
and Law texture features, predominantly from the enhancing tumor region in FLAIR
sequenced images.

CONCLUSIONS: Our
preliminary results suggest that radiomic features, particularly in the
necrotic region of the tumor may provide complementary diagnostic information
on routine MR imaging sequences that may improve the distinction between the
tumors with activated and wild type AKT pathways in patients with GBM.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Introduction

Glioblastoma (GBM) is the most
common and aggressive form of cancer that begins in the brain. The disease is
difficult to detect without MR imaging due to the non-specific nature of the
symptoms, which begin mildly and can progress rapidly as the disease
progresses.1  Without treatment, survival is typically 3
months.  With treatment, survival
increases to between 12 and 15 months, with only 3 to 5 % surviving past 5
years past diagnosis. 

The AKT/PI3K signaling pathway is
involved in cellular proliferation, and in many cancers, this pathway is
overactive. PTEN acts as an antagonist to the AKT pathway by dephosphorylating
PIP3 to PIP2, which inhibits AKT’s ability to bind to the cell membrane,
decreasing cellular proliferation.2 PTEN is the most commonly
disrupted tumor suppressor, with mutations found in 20 – 40% of GBM tumors. PTEN
is activated by phosphatidylinositol 3-kinase (PI3K), a product of PIK3CA3. If either PTEN or PIK3CA
have a mutation resulting in a loss of function the AKT pathway is activated,
which can cause therapy resistance and tumorigenesis1. Mutational status is
currently determined through DNA sequencing of surgically resected specimens or
biopsy samples. Unfortunately, this requires invasive intervention, as well as
being prone to sampling errors (as gene profiling is assayed only on a small portion
of tissue).

GBM is routinely diagnosed using
Magnetic Resonance Imaging (MRI). While MRI provides structural and functional
non-invasive characterization of the tumor, it is not currently effective in
capturing the underlying tumor pathology, and molecular heterogeneity. To
definitively diagnose GBM, a stereotactic biopsy or craniotomy with pathologic
confirmation is required. Biopsy provides valuable genetic information, but it
is invasive and does not capture the heterogeneity of the tumor.

Radiomics is the use of
data-characterization algorithms to extract large sums of quantitative textural
and shape information from medical images. While radiomics is applicable to
many medical conditions, it is the most developed and commonly applied for
oncology.4 Radiogenomics is the term
used to refer to finding the correlation between cancer imaging features and
genetic expression information5. This approach has had some
previous success in determining the associated genetics with MRI phenotypes in
GBM.6 Other studies have found that
volumetric MRI features are significantly predictive of mutation status in TP53,
RB1, NF1, EGFR, and PDGFRA in GBM cases. We are not aware of previous studies
associating the AKT pathway with MRI features in GBM.

The purpose of this study is to
determine if radiogenomic analysis of the routinely generated MR images can
give insight to the AKT pathway status in cases of GBM. We wish to assess
whether these radiomic features present themselves differently within the
different regions of GBM (enhancing tumor, edema, necrosis) and across 3
routine multiparametric MR images (Gd-T1WI, T2WI, FLAIR). In this retrospective
study, we identified radiomic features in the necrotic region in FLAIR and T2
images that had significant correlation with AKT pathway status among a cohort
of 58 GBM patients. The eventual goal of this study is to develop a tool that
is more readily available world-wide than biopsy and genetic sequencing to
diagnose genetic variants in glioblastoma. Genetic sequencing is also expensive
and not readily available to all GBM patients around the world.  In this study, we find radiomic features that
are the most correlated with AKT pathway mutational status, and assess the
predictive value of those features using a random forest analysis.

MATERIALS AND METHODS

Study Population

The cohort for this study consists
of 75 retrospectively analyzed pre-treatment multi-parametric MRI scans from
The Cancer Imaging Archive (TCIA). The set included 30 individuals with an
activated AKT pathway, and 45 with a wildtype AKT pathway (Table 1). Inclusion
criteria in the training cohort were: (1) the availability of all three of the
routine MRI sequences (T1, T2, FLAIR) for patients, pre-treatment, in the
training cohort, (2) MRI scans with diagnostic image quality, and (3) the
availability of AKT pathway status, including mutational data for PTEN and
PIK3CA.

Table 1,
Distribution of age, gender and race amongst cases with wild type and activated
AKT pathways. p-values were generated using Chi-squared tests.

Stratified by AKT
Pathway Status

Wild Type

Activated

p

n

45

30

 

age (mean(sd))

58.07(13.21)

62.43(12.62)

0.158

gender = male (%)

29(64.4)

16(53.3)

0.47

race (%)

 

0.659

   
Asian

1(2.2)

0(0.0)

   
Black or African American

2(4.4)

2(6.7)

   
White

42(93.3)

28(93.3)

 

 

Image Registration

For every patient
scan, 3D Slicer 4.5 was used to co-register T2w and FLAIR with reference to Gd Gadolinium
(GD)-T1w using 3D affine registration with 12 degrees of freedom encoding
rotation. Every MRI slice was resampled to a uniform pixel spacing of 0.5 x 0.5
mm2 and interpolated to a 1 mm slice thickness to account for
resolution variability. Annotation of each 2-dimensional slice with visible
tumor was performed by an expert reader in 3 sections: edema, enhancing tumor,
and tumor necrosis. Tumor necrosis was represented by hypo-intense regions
generally located centrally in the tumoral region in Gd-T1w sequences.
Hyper-intense regions in FLAIR sequences correlate with greater interstitial
leakage and lower cell density, signifying edema.

3D Radiomic Feature Extraction

For each
patient, 13 first order statistics, 13 second order statistics, 39 Haralick
features (captures tumor heterogeneity), 96 Gabor features (captures structural
detail at different scales and orientations) , 152 Law’s energy features
(captures presence of spots, edges, waves and ripples in an image), 13 CoLlAGe7 features, and 38 shape
features were extracted in the context of three dimensions for each compartment
(tumor necrosis, enhancing tumor, edema) on a per-voxel basis for each MRI
sequence (FLAIR, Gd-T1w, T2w). All feature calculations were performed using in-house
software implemented in MATLAB R 2016b platform (Mathworks, Natick, MA).

AKT Pathway Activation
Classification

               The
mutation status of PIK3CA and PTEN were determined for each patient using data
from TumorPortal.8 Mutation status was
determined through DNA sequencing. Patients with a tumor that had a PTEN
inactivating mutation, or a PIK3CA activating mutation were classified as having
an activated AKT pathway.

Feature Selection and Statistical
Analysis

The cohort was
divided into a random set of 37 cases to train the model and 38 to test. The
standard deviation and median were computed for each feature, and these
statistics were normalized and compared between patients with wild type AKT
pathway status and activated AKT pathway status. Feature statistics were ranked
by computing the mutual information between each feature and AKT pathway
status, employing three-fold cross validation. The five features that
consistently had the most mutual information were used in a Quadratic
Discriminant Analysis (QDA) classification. 
The classifier was then applied to the independent test set and assessed
for sensitivity and specificity in predicting AKT pathway status.

RESULTS

The mutual
information analysis reported that the five features that shared the most
information with AKT status activation were one Gabor feature from the necrotic
region in the T1 sequence, in addition to two intensity features and two
Haralick features from the enhancing tumor region in the FLAIR sequence (Table
2). In the training set, the QDA classifier had an accuracy
rate of 0.7329. The overall prediction accuracy for the testing set was 0.7368
(Sensitivity = 0.90, Specificity = 0.844).

Table 2:
Table of the five  features with the
highest mutual information with AKT pathway status, and the sequence and
segment that those features were found.

Feature Name

Sequence

Region

Feature Type

median-XY
Orient=0.5236, XZ Orient=0, Bandwidth=1, Wavelength=1.4142

T1

Necrosis

Gabor

median-raw_intensity

FLAIR 

Enhancing Tumor

Intensity

std-raw_intensity

FLAIR

Enhancing Tumor

Intensity

median-mean ws=5

FLAIR

Enhancing Tumor

Haralick

median-median ws=5

FLAIR

Enhancing Tumor

Haralick

 

 

DISCUSSION

We found that
texture features in the enhancing tumor region of a GBM tumor could be
correlated with whether the AKT pathway is activated or wildtype. Contrast enhancement
in MRI imaging is indicative of the localized deterioration of the blood-brain
barrier, which is generally associated with high-grade gliomas. 9 Previous studies have found
that the AKT pathway influences the vascular endothelial growth factor (VEGF)
which affects the permeability of the blood brain barrier, so it is not surprising
that the AKT pathway may be associated with the enhancing tumor region in some
way. 2  This analysis was limited in sample size. With
more patients, feature selection could be more thorough and classification
could be more accurate, so this is something to explore in future work. The AKT
pathway activation status was characterized using the mutation of PTEN and/or
PIK3CA as markers. This characterization method is not perfect and could allow
for some misclassification of activated and wildtype individuals that may skew
results.    

 

CONCLUSIONS

In this study,
we investigated the distinction in AKT pathway status between 3D MRI features
in GBM across the regions (edema, enhancing tumor, necrosis) of the tumor. Our
results suggest that radiomic analysis on routinely acquired MR imaging may
enable discrimination between wildtype and activated AKT pathways in the enhancing
tumor region of GBM, specifically intensity and Haralick features in the FLAIR
sequenced images.

 

REFERENCES

 

1.            Li, X. et al.
PI3K/Akt/mTOR signaling pathway and targeted therapy for glioblastoma. Oncotarget
7, 33440–33450 (2016).

2.            Kilic,
E. et al. The phosphatidylinositol-3 kinase/Akt pathway mediates VEGF’s
neuroprotective activity and induces blood brain barrier permeability after
focal cerebral ischemia. FASEB J. Off. Publ. Fed. Am. Soc. Exp. Biol. 20,
1185–1187 (2006).

3.            Samuels,
Y. & Waldman, T. Oncogenic Mutations of PIK3CA in Human Cancers. Curr.
Top. Microbiol. Immunol. 347, 21–41 (2010).

4.            Gillies,
R. J., Kinahan, P. E. & Hricak, H. Radiomics: Images Are More than Pictures,
They Are Data. Radiology 278, 563–577 (2015).

5.            Mazurowski,
M. A. Radiogenomics: what it is and why it is important. J. Am. Coll.
Radiol. JACR 12, 862–866 (2015).

6.            Gevaert,
O. et al. Glioblastoma multiforme: exploratory radiogenomic analysis by
using quantitative image features. Radiology 273, 168–174 (2014).

7.            Prasanna,
P., Tiwari, P. & Madabhushi, A. Co-occurrence of Local Anisotropic Gradient
Orientations (CoLlAGe): A new radiomics descriptor. Sci. Rep. 6,
37241 (2016).

8.            Lawrence,
M. S. et al. Discovery and saturation analysis of cancer genes across 21
tumor types. Nature 505, 495–501 (2014).

9.            Mabray,
M. C., Barajas, R. F. & Cha, S. Modern Brain Tumor Imaging. Brain Tumor
Res. Treat. 3, 8–23 (2015).

ABSTRACT

BACKGROUND AND
PURPOSE: Activation of the AKT pathway has a significant role in cellular
proliferation, migration, and apoptosis, and is associated with poor prognosis
in GBM. Biopsy and genetic sequencing are used to diagnose this mutation, which
is not a routine procedure in the majority of hospitals and clinics. In this
work, we explored the feasibility of directional gradients (Gabor) and local
intensity statistics (Haralick, Law) texture features obtained from different
tumor-specific sub-compartments (enhancing, non-enhancing, infiltrating edges,
and necrotic regions) on T1, T2 and FLAIR sequence in non-invasively predicting
AKT pathway mutational status from routinely acquired MRI.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

 MATERIALS AND METHODS: A retrospective study of brain tumor MR
imaging performed on images from TCIA. 75 patient studies were analyzed. Brain
lesions on MR imaging were manually annotated by an expert neuroradiologist. A
set of three dimensional radiomic features was extracted for every lesion on
each MR imaging sequence: gadolinium T1WI, T2WI, and FLAIR, and within each
compartment: enhancing tumor, edema, necrosis. Rank-sum tests were used to find
features that correlated with the AKT pathway status of the patient studies.

RESULTS: 5
radiomic texture features with the most mutual information were Gabor, Sobel,
and Law texture features, predominantly from the enhancing tumor region in FLAIR
sequenced images.

CONCLUSIONS: Our
preliminary results suggest that radiomic features, particularly in the
necrotic region of the tumor may provide complementary diagnostic information
on routine MR imaging sequences that may improve the distinction between the
tumors with activated and wild type AKT pathways in patients with GBM.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Introduction

Glioblastoma (GBM) is the most
common and aggressive form of cancer that begins in the brain. The disease is
difficult to detect without MR imaging due to the non-specific nature of the
symptoms, which begin mildly and can progress rapidly as the disease
progresses.1  Without treatment, survival is typically 3
months.  With treatment, survival
increases to between 12 and 15 months, with only 3 to 5 % surviving past 5
years past diagnosis. 

The AKT/PI3K signaling pathway is
involved in cellular proliferation, and in many cancers, this pathway is
overactive. PTEN acts as an antagonist to the AKT pathway by dephosphorylating
PIP3 to PIP2, which inhibits AKT’s ability to bind to the cell membrane,
decreasing cellular proliferation.2 PTEN is the most commonly
disrupted tumor suppressor, with mutations found in 20 – 40% of GBM tumors. PTEN
is activated by phosphatidylinositol 3-kinase (PI3K), a product of PIK3CA3. If either PTEN or PIK3CA
have a mutation resulting in a loss of function the AKT pathway is activated,
which can cause therapy resistance and tumorigenesis1. Mutational status is
currently determined through DNA sequencing of surgically resected specimens or
biopsy samples. Unfortunately, this requires invasive intervention, as well as
being prone to sampling errors (as gene profiling is assayed only on a small portion
of tissue).

GBM is routinely diagnosed using
Magnetic Resonance Imaging (MRI). While MRI provides structural and functional
non-invasive characterization of the tumor, it is not currently effective in
capturing the underlying tumor pathology, and molecular heterogeneity. To
definitively diagnose GBM, a stereotactic biopsy or craniotomy with pathologic
confirmation is required. Biopsy provides valuable genetic information, but it
is invasive and does not capture the heterogeneity of the tumor.

Radiomics is the use of
data-characterization algorithms to extract large sums of quantitative textural
and shape information from medical images. While radiomics is applicable to
many medical conditions, it is the most developed and commonly applied for
oncology.4 Radiogenomics is the term
used to refer to finding the correlation between cancer imaging features and
genetic expression information5. This approach has had some
previous success in determining the associated genetics with MRI phenotypes in
GBM.6 Other studies have found that
volumetric MRI features are significantly predictive of mutation status in TP53,
RB1, NF1, EGFR, and PDGFRA in GBM cases. We are not aware of previous studies
associating the AKT pathway with MRI features in GBM.

The purpose of this study is to
determine if radiogenomic analysis of the routinely generated MR images can
give insight to the AKT pathway status in cases of GBM. We wish to assess
whether these radiomic features present themselves differently within the
different regions of GBM (enhancing tumor, edema, necrosis) and across 3
routine multiparametric MR images (Gd-T1WI, T2WI, FLAIR). In this retrospective
study, we identified radiomic features in the necrotic region in FLAIR and T2
images that had significant correlation with AKT pathway status among a cohort
of 58 GBM patients. The eventual goal of this study is to develop a tool that
is more readily available world-wide than biopsy and genetic sequencing to
diagnose genetic variants in glioblastoma. Genetic sequencing is also expensive
and not readily available to all GBM patients around the world.  In this study, we find radiomic features that
are the most correlated with AKT pathway mutational status, and assess the
predictive value of those features using a random forest analysis.

MATERIALS AND METHODS

Study Population

The cohort for this study consists
of 75 retrospectively analyzed pre-treatment multi-parametric MRI scans from
The Cancer Imaging Archive (TCIA). The set included 30 individuals with an
activated AKT pathway, and 45 with a wildtype AKT pathway (Table 1). Inclusion
criteria in the training cohort were: (1) the availability of all three of the
routine MRI sequences (T1, T2, FLAIR) for patients, pre-treatment, in the
training cohort, (2) MRI scans with diagnostic image quality, and (3) the
availability of AKT pathway status, including mutational data for PTEN and
PIK3CA.

Table 1,
Distribution of age, gender and race amongst cases with wild type and activated
AKT pathways. p-values were generated using Chi-squared tests.

Stratified by AKT
Pathway Status

Wild Type

Activated

p

n

45

30

 

age (mean(sd))

58.07(13.21)

62.43(12.62)

0.158

gender = male (%)

29(64.4)

16(53.3)

0.47

race (%)

 

0.659

   
Asian

1(2.2)

0(0.0)

   
Black or African American

2(4.4)

2(6.7)

   
White

42(93.3)

28(93.3)

 

 

Image Registration

For every patient
scan, 3D Slicer 4.5 was used to co-register T2w and FLAIR with reference to Gd Gadolinium
(GD)-T1w using 3D affine registration with 12 degrees of freedom encoding
rotation. Every MRI slice was resampled to a uniform pixel spacing of 0.5 x 0.5
mm2 and interpolated to a 1 mm slice thickness to account for
resolution variability. Annotation of each 2-dimensional slice with visible
tumor was performed by an expert reader in 3 sections: edema, enhancing tumor,
and tumor necrosis. Tumor necrosis was represented by hypo-intense regions
generally located centrally in the tumoral region in Gd-T1w sequences.
Hyper-intense regions in FLAIR sequences correlate with greater interstitial
leakage and lower cell density, signifying edema.

3D Radiomic Feature Extraction

For each
patient, 13 first order statistics, 13 second order statistics, 39 Haralick
features (captures tumor heterogeneity), 96 Gabor features (captures structural
detail at different scales and orientations) , 152 Law’s energy features
(captures presence of spots, edges, waves and ripples in an image), 13 CoLlAGe7 features, and 38 shape
features were extracted in the context of three dimensions for each compartment
(tumor necrosis, enhancing tumor, edema) on a per-voxel basis for each MRI
sequence (FLAIR, Gd-T1w, T2w). All feature calculations were performed using in-house
software implemented in MATLAB R 2016b platform (Mathworks, Natick, MA).

AKT Pathway Activation
Classification

               The
mutation status of PIK3CA and PTEN were determined for each patient using data
from TumorPortal.8 Mutation status was
determined through DNA sequencing. Patients with a tumor that had a PTEN
inactivating mutation, or a PIK3CA activating mutation were classified as having
an activated AKT pathway.

Feature Selection and Statistical
Analysis

The cohort was
divided into a random set of 37 cases to train the model and 38 to test. The
standard deviation and median were computed for each feature, and these
statistics were normalized and compared between patients with wild type AKT
pathway status and activated AKT pathway status. Feature statistics were ranked
by computing the mutual information between each feature and AKT pathway
status, employing three-fold cross validation. The five features that
consistently had the most mutual information were used in a Quadratic
Discriminant Analysis (QDA) classification. 
The classifier was then applied to the independent test set and assessed
for sensitivity and specificity in predicting AKT pathway status.

RESULTS

The mutual
information analysis reported that the five features that shared the most
information with AKT status activation were one Gabor feature from the necrotic
region in the T1 sequence, in addition to two intensity features and two
Haralick features from the enhancing tumor region in the FLAIR sequence (Table
2). In the training set, the QDA classifier had an accuracy
rate of 0.7329. The overall prediction accuracy for the testing set was 0.7368
(Sensitivity = 0.90, Specificity = 0.844).

Table 2:
Table of the five  features with the
highest mutual information with AKT pathway status, and the sequence and
segment that those features were found.

Feature Name

Sequence

Region

Feature Type

median-XY
Orient=0.5236, XZ Orient=0, Bandwidth=1, Wavelength=1.4142

T1

Necrosis

Gabor

median-raw_intensity

FLAIR 

Enhancing Tumor

Intensity

std-raw_intensity

FLAIR

Enhancing Tumor

Intensity

median-mean ws=5

FLAIR

Enhancing Tumor

Haralick

median-median ws=5

FLAIR

Enhancing Tumor

Haralick

 

 

DISCUSSION

We found that
texture features in the enhancing tumor region of a GBM tumor could be
correlated with whether the AKT pathway is activated or wildtype. Contrast enhancement
in MRI imaging is indicative of the localized deterioration of the blood-brain
barrier, which is generally associated with high-grade gliomas. 9 Previous studies have found
that the AKT pathway influences the vascular endothelial growth factor (VEGF)
which affects the permeability of the blood brain barrier, so it is not surprising
that the AKT pathway may be associated with the enhancing tumor region in some
way. 2  This analysis was limited in sample size. With
more patients, feature selection could be more thorough and classification
could be more accurate, so this is something to explore in future work. The AKT
pathway activation status was characterized using the mutation of PTEN and/or
PIK3CA as markers. This characterization method is not perfect and could allow
for some misclassification of activated and wildtype individuals that may skew
results.    

 

CONCLUSIONS

In this study,
we investigated the distinction in AKT pathway status between 3D MRI features
in GBM across the regions (edema, enhancing tumor, necrosis) of the tumor. Our
results suggest that radiomic analysis on routinely acquired MR imaging may
enable discrimination between wildtype and activated AKT pathways in the enhancing
tumor region of GBM, specifically intensity and Haralick features in the FLAIR
sequenced images.

 

REFERENCES

 

1.            Li, X. et al.
PI3K/Akt/mTOR signaling pathway and targeted therapy for glioblastoma. Oncotarget
7, 33440–33450 (2016).

2.            Kilic,
E. et al. The phosphatidylinositol-3 kinase/Akt pathway mediates VEGF’s
neuroprotective activity and induces blood brain barrier permeability after
focal cerebral ischemia. FASEB J. Off. Publ. Fed. Am. Soc. Exp. Biol. 20,
1185–1187 (2006).

3.            Samuels,
Y. & Waldman, T. Oncogenic Mutations of PIK3CA in Human Cancers. Curr.
Top. Microbiol. Immunol. 347, 21–41 (2010).

4.            Gillies,
R. J., Kinahan, P. E. & Hricak, H. Radiomics: Images Are More than Pictures,
They Are Data. Radiology 278, 563–577 (2015).

5.            Mazurowski,
M. A. Radiogenomics: what it is and why it is important. J. Am. Coll.
Radiol. JACR 12, 862–866 (2015).

6.            Gevaert,
O. et al. Glioblastoma multiforme: exploratory radiogenomic analysis by
using quantitative image features. Radiology 273, 168–174 (2014).

7.            Prasanna,
P., Tiwari, P. & Madabhushi, A. Co-occurrence of Local Anisotropic Gradient
Orientations (CoLlAGe): A new radiomics descriptor. Sci. Rep. 6,
37241 (2016).

8.            Lawrence,
M. S. et al. Discovery and saturation analysis of cancer genes across 21
tumor types. Nature 505, 495–501 (2014).

9.            Mabray,
M. C., Barajas, R. F. & Cha, S. Modern Brain Tumor Imaging. Brain Tumor
Res. Treat. 3, 8–23 (2015).

x

Hi!
I'm Mary!

Would you like to get a custom essay? How about receiving a customized one?

Check it out