- Email: [email protected]
Computed Tomography Radiomics Predicts HPV Status and Local Tumor Control After Definitive Radiochemotherapy in Head and Neck Squamous Cell Carcinoma
Accepted Manuscript CT radiomics predicts HPV status and local tumor control after definitive radiochemotherapy in head and neck squamous cell carcinoma Marta Bogowicz, MSc, Oliver Riesterer, MD, Kristian Ikenberg, MD, Sonja Stieb, MD, Holger Moch, MD, Prof, Gabriela Studer, MD, Prof, Matthias Guckenberger, MD, Prof, Stephanie Tanadini-Lang, PhD PII:
S0360-3016(17)31011-8
DOI:
10.1016/j.ijrobp.2017.06.002
Reference:
ROB 24297
To appear in:
International Journal of Radiation Oncology • Biology • Physics
Received Date: 22 December 2016 Revised Date:
15 May 2017
Accepted Date: 5 June 2017
Please cite this article as: Bogowicz M, Riesterer O, Ikenberg K, Stieb S, Moch H, Studer G, Guckenberger M, Tanadini-Lang S, CT radiomics predicts HPV status and local tumor control after definitive radiochemotherapy in head and neck squamous cell carcinoma, International Journal of Radiation Oncology • Biology • Physics (2017), doi: 10.1016/j.ijrobp.2017.06.002. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
CT radiomics predicts HPV status and local tumor control after definitive radiochemotherapy in head and neck squamous cell carcinoma
Marta Bogowicz, MSc, 1Oliver Riesterer, MD, 2Kristian Ikenberg, MD, 1Sonja Stieb, MD, 2Holger
RI PT
1
Moch, MD, Prof, 1Gabriela Studer, MD, Prof, 1Matthias Guckenberger, MD, Prof, 1Stephanie TanadiniLang, PhD
Department of Radiation Oncology and 2Department of Pathology and Molecular Pathology, University
Hospital Zurich, University of Zurich, Switzerland
Address: Department of Radiation Oncology University Hospital Zurich Rämistrasse 100
Switzerland
Corresponding author: Marta Bogowicz
TE D
8091 Zurich
M AN U
Running title: Radiomics predicts HPV and tumor control
SC
1
EP
Email: [email protected] Phone: +41 44 255 4116
AC C
Fax: +41 44 255 4547
Conflict of interest: Authors declare no conflict of interest.
Acknowledgments:
The project was supported by the Clinical Research Priority Program Tumor Oxygenation of the University of Zurich, by a grant of the Matching Fund of the University of Zurich and by a research grant from Merck (Schweiz) AG.
ACCEPTED MANUSCRIPT
Summary: This study shows that higher pre-treatment tumor heterogeneity, quantified using radiomics of contrast-enhanced CT imaging, is associated with worse prognosis in head and neck squamous
RI PT
cell carcinoma and the absence of HPV infection. Overall, our analysis provides an evidence that a radiomic signature might be a useful input in risk assessment in addition to clinical parameters
AC C
EP
TE D
M AN U
SC
(tumor stage, tumor volume and HPV status).
ACCEPTED MANUSCRIPT CT radiomics predicts HPV status and local tumor control after definitive radiochemotherapy in head and neck squamous cell carcinoma
RI PT
Running title: Radiomics predicts HPV and tumor control
Abstract Purpose
SC
This study aimed to predict local tumor control (LC) after radiochemotherapy of head and neck squamous cell carcinoma (HNSCC) and HPV status using CT radiomics.
M AN U
Materials and methods
HNSCC patients treated with definitive radiochemotherapy were included in the retrospective study approved by local ethical commission (93 and 56 patients in the training and validation cohorts, respectively). 317 CT radiomic features, including shape-, intensity-, texture- and transform-based features, were calculated in the primary tumor region. A Cox and logistic
TE D
regression models were built to predict LC and HPV status, respectively. The best performing features in the univariate analysis were included in the multivariate analysis, after exclusion
EP
of redundant features. The quality of the models was assessed using the concordance index (CI) for modelling of LC and receiver operating characteristics area under the curve (AUC)
AC C
for HPV status prediction. The radiomics LC model was compared to model incorporating clinical parameters (tumor stage, volume and HPV status) and a mixed model. Results
A radiomic signature, comprising three features, was significantly associated with LC (CItraining=0.75 and CIvalidation=0.78), showing that tumors with a more heterogeneous CT density distribution are at risk for decreased LC. Addition of clinical parameters to the radiomics model slightly improved the model in the training but not in validation cohort. Another radiomic signature showed a good performance in HPV status prediction
ACCEPTED MANUSCRIPT (AUCtraining=0.85 and AUCvalidation=0.78) and indicated that HPV positive tumors have more homogenous CT density distribution. Conclusion HNSCC tumor density heterogeneity, quantified by CT radiomics, is associated with LC after
RI PT
radiochemotherapy and HPV status.
SC
Introduction
Head and neck squamous cell carcinoma (HNSCC) is the fifth most common cancer
M AN U
worldwide and the standard of care for advanced unresectable cases is definitive radiochemotherapy. Clinical outcome after radiochemotherapy is very heterogeneous ranging from about 80% locoregional control and 5 years overall survival in HPV-positive oropharyngeal carcinoma (OPC) to below 50% in HPV-negative OPC and non-oropharyngeal
TE D
cancers [1, 2]. In all cases distant metastasis is infrequent and mainly a consequence of locoregional treatment failure. Therefore, improvement of radiochemotherapy is essential. An important approach is the development of tumor biomarkers that can predict locoregional
EP
tumor control in order to identify patients for treatment intensification. Radiomics is a promising tool for non-invasive characterization of tumor phenotype. It
AC C
extracts a large number of quantitative features from radiographic images: regions of interest are characterized regarding their shape, intensities and texture of both spatial and frequency domains (for example using a wavelet transform) [3]. A correlation between radiomic features and treatment response has been shown using different image modalities and for different tumor types [4, 5]. To provide a biological interpretation of radiomics several groups tried to relate a radiomic signature (set of radiomic parameters) to differential gene expressions including tumor proliferation and molecular subtypes [6, 7].
ACCEPTED MANUSCRIPT Although several CT radiomics-based prognostic models have been proposed for HNSCC [8], it is not fully understood why a particular radiomic phenotype is associated with treatment response. CT radiomics of the primary tumor was successfully used to stratify patients into low and high risk groups in terms of patient overall survival [6, 9, 10, 8]. It was shown to
RI PT
further improve survival models based on clinical parameters, such as TNM stage and tumor volume [6]. While overall survival is an important clinical endpoint, it is also influenced by non-tumor related factors [11], which cannot be captured by pretreatment radiomics of the
SC
primary tumor. However, especially in the case of HNSCC, it is important to understand the influence of intra-tumor heterogeneity on the local response to treatment. Therefore, a study
M AN U
using local tumor control as endpoint is of high clinical interest. Furthermore, the correlation of radiomic phenotype to the important outcome prognostic factor HPV is of interest and links image analysis with tumor biology [12, 13]. Exploratory studies have already indicated a relationship between HPV infection and the heterogeneity of imaging-based tumor density
TE D
[12, 13]. However, these studies were based on small patient numbers and lacked validation cohorts.
In this study we investigated, if CT-based radiomic phenotype of HNSCC is correlated with
EP
local tumor control after radiochemotherapy. Furthermore, using two big patient cohorts we
AC C
studied the influence of HPV infection on tumor radiomics.
Materials and methods
Studied population
This study was a retrospective analysis approved by a local ethical commission. The training cohort consisted of 93 stage III and IV HNSCC patients treated between 2003-2013 with definitive radiotherapy (on average 70 Gy) either in combination with cisplatin (40 mg/m2, up
ACCEPTED MANUSCRIPT to 7 cycles) or cetuximab (loading dose 400 mg/m2 followed by 250 mg/m2 weekly). Induction chemotherapy and/or prior surgery (biopsy allowed) were exclusion criteria. Additionally, data from 56 patients with the same tumor stages from an institutional phase III prospective study (NCT01435252) with a standardized planning CT imaging protocol were
RI PT
selected as validation cohort. In the latter study, patients were treated with triple therapy consisting of radiotherapy and weekly concurrent cisplatin/cetuximab (same doses as in training cohort) with or without consolidation cetuximab (500 mg/m2 biweekly x 6).
SC
Radiotherapy was delivered using intensity modulated technique in all patients.
All patients underwent pretreatment contrast-enhanced planning CT imaging. Patient
M AN U
characteristics and imaging protocols are presented in the Table 1. In all patients the HPV status was determined by immunohistochemical p16 staining. Local recurrence was defined as biopsy proven recurrence within the high dose PTV of the primary tumor.
Follow-up
TE D
consisted of 2-3 monthly examinations by ENT specialists and repetitive FDG-PET/CTs.
Radiomics analysis and image post-processing
EP
An in-house developed radiomics software implementation written in Python programing language was used. It provides 3D analysis of medical images using all four feature extraction
• • •
AC C
methods:
shape (n = 11)
intensity (n = 5)
texture: the Gray Level Co-occurrence Matrix (n = 14), the Neighborhood Gray Tone Difference Matrix (n = 4) and the Gray Level Size Zone Matrix (n = 11)
•
wavelet transform using Coiflet function (8 sub-bands, n = 272).
In total 317 features were extracted per patient. For intensity, texture and wavelet analysis images were resized to the symmetrical voxels of size of 3.3 mm, corresponding to the lowest
ACCEPTED MANUSCRIPT image resolution (sagittal) in the training cohort. Additionally, Hounsfield Units range of -20 to 180 HU was applied to limit the analysis to soft tissue only. To quantify the texture images were resampled to equally spaced bins of 5 HU. A detailed description of radiomics implementation is presented in the Appendix.
RI PT
The tumor heterogeneity was studied inside the primary GTV defined on the contrastenhanced planning CT by a radiation oncologist with more than 10 years of experience. Contours were post-processed to account for metal artifacts occurring in the CT images.
SC
Contours were removed from artifact-affected slices. Scans with more than a half of the
Radiomic inter-features correlations
M AN U
contoured slices affected by metal artifacts were not included in the analysis.
Radiomic features were grouped according to principal component analysis using data from
TE D
the training cohort to account for the inter-features correlations. The Horn method was used to choose the optimal number of retained components [14]. The contribution of radiomic features to principal components were determined. Features were assigned to a principal
EP
component to which they contributed the most. Each principal component corresponded to
AC C
one group of correlated features. The statistical analysis was performed in R (v. 3.2.4) [15].
Radiomics prognostic value of local tumor control
Radiomic features prognostic for local tumor control were preselected in the univariate Cox regression analysis. The results were corrected for multiple testing with false discovery rate (FDR) < 0.20. Additionally, the correlated features were excluded from the further analysis by selecting only the most prognostic parameter per principal component feature group. Feature prognostic value was quantified using concordance index (CI) [16]. The preselected features
ACCEPTED MANUSCRIPT were entered in the multivariate Cox regression analysis. The radiomics-based model was compared with a multivariate model created with clinical parameters: tumor volume, T stage (T1/T2 vs T3/T4) and HPV status. Finally, a mixed model was created, where both radiomic features and clinical parameters were incorporated in the multivariate Cox model. All
RI PT
multivariate models were built using Akaike information criterion in the backward selection of variables. The prognostic value of obtained models was compared based on the CI, using Wilcoxon test (p-value < 0.05) and the bootstrap method with 100 randomly selected samples
SC
to calculate the CI distribution. The results were verified in the validation cohort. Curves were split based on the optimal threshold from the receiver operating characteristic curve at 18
M AN U
months maximizing both sensitivity and specificity and compared using G-rho test (pvalue < 0.05).
TE D
Prediction of HPV status
The radiomic features for HPV status prediction were also preselected using principal component analysis in combination with univariate logistic regression. The univariate model
EP
accuracy was quantified by the area under receiver operating characteristic curve (AUC). The results were corrected for multiple testing with FDR < 0.20. A radiomic feature characterized
AC C
by the highest AUC was selected in each principal component group. This set was later used in multivariate logistic regression with backward selection of the variables. The cutoff threshold of the final model was set to maximize both sensitivity and specificity in the training cohort. The results were again verified in the validation cohort.
ACCEPTED MANUSCRIPT
Results
RI PT
Radiomics-based prognostic model for local tumor control
Nine components were selected in the principal component analysis, which lead to nine groups of correlated radiomic features. The group sizes varied from 21 to 66 features. In the
SC
univariate Cox regression analysis 83 out of 317 radiomic features were found to be prognostic for local tumor control (FDR < 0.20). The list of the most prognostic features,
M AN U
accounting for inter-feature correlations, is shown in Table 2. In the multivariate analysis with the backward selection of variables three (HHH large zone high gray-level emphasis, LLL sum entropy, and LLH difference variance) out of nine previously mentioned features were selected, resulting in a model with CI of 0.75. This model was found to be prognostic also in
TE D
the validation cohort with CI = 0.78 and significant for stratification into groups with low and high risk of recurrence (Figure 1; Table 3). Additionally, the model was prognostic in the subgroup analysis of oropharyngeal carcinoma as well as HPV negative tumors (Appendix
EP
Table 1S and Figure 2S). Different CT scanners and kV settings did not show a major
AC C
influence on the radiomic features included in the model (Appendix Figure 6S).
Comparison of the radiomics-based and clinical-based prognostic model for local tumor control
The radiomics-based model was compared with a model incorporating known prognostic clinical parameters (tumor volume, T stage and HPV status) as well as with a mixed-model (radiomics and clinical parameters). The parameters chosen in the multivariate analysis and the results of Wilcoxon test, studying differences between the models, are presented in the
ACCEPTED MANUSCRIPT Table 3. The combination of radiomics (LLH difference variance) and clinical parameters (HPV and tumor volume) was found to have the highest prognostic power (CI = 0.80) in the training cohort, however performed worse than the radiomics-based model in the validation cohort. The same trend was observed for the subgroup of oropharyngeal carcinoma (Appendix
RI PT
Table 1S). In the mixed model the single radiomic feature was not significant however, it had a significant impact on the improvement of the model discriminative power (Table 2S).
SC
Prediction of HPV status
M AN U
In total, 86 radiomic features were found to be predictive for HPV status (FDR < 0.20). Per group of correlated features, the one characterized with the highest AUC was chosen (Table 4) and entered in the multivariate logistic regression. The final model comprised radiomic features: LLL standard deviation
•
LLL small zone high gray-level emphasis
•
HHL difference entropy
•
coefficient of variation
EP
TE D
•
four
AC C
The final radiomic model for HPV status prediction showed a good performance in both training (AUC = 0.85) and validation cohort (AUC = 0.78). The specificity of the model was higher in the validation in comparison to the training cohort (specificitytraining = 0.72 and specificityvalidation = 0.82), whereas the sensitivity was lower (sensitivitytraining = 0.72 and sensitivitytraining = 0.62). Figure 2 presents the receiver operating characteristics for the model. The HPV radiomic signature was independent from the local control signature as radiomic features in these two models were not correlated and the HPV prediction model was valid in patient subgroups regardless of the tumor control (Appendix Figure 3S). Similar to the local
ACCEPTED MANUSCRIPT tumor control model, we have not observed major variations in our HPV radiomic signature depending on the CT scanner and kV settings (Appendix Figure 7S). Discussion This study evaluated the prognostic value of contrast enhanced CT-based radiomics in head
RI PT
and neck squamous cell carcinoma regarding local tumor control and HPV status. The radiomics-based local control model was compared to a model based on prognostic clinical parameters (tumor stage, tumor volume and HPV status) and to the mixed model (clinical plus
SC
radiomics). The performance of the models was verified in an independent cohort of patients. Although the radiomics-based model achieved lower concordance index in comparison to the
M AN U
mixed model in the training cohort, it showed a significantly higher performance in the validation cohort. The stable performance of radiomics-based model in both training and validation patient cohorts stresses its high credibility. Radiomics was also shown to be closely associated with the HPV status with AUC values of 0.85 and 0.78 in the training and the
TE D
validation cohort, respectively.
In the case of HNSCC, locoregional control closely correlates with overall survival [17]. Prediction of local treatment efficacy is of high interest in order to modify treatment
EP
according to the individual tumor’s radiosensitivity. The radiomics model created by us was highly prognostic for local tumor control and addition of the clinically relevant biomarker,
AC C
HPV status, could only slightly improve the model in the training cohort but not in the validation cohort. Similar results were observed for the analysis of OPC patients only (Appendix Table 1S), which indicates slight overfitting of the mixed model (radiomics and clinical predictors) in the training cohort. The value of the radiomic feature (LLH difference variance) in the mixed model should be studied in more details. This covariate was not significant in the multivariate modeling however, had a positive impact on the discriminative power of the model, especially in the validation cohort. The training and validation cohorts were not perfectly matched. The validation cohort was characterized by higher percentage of
ACCEPTED MANUSCRIPT more advanced tumors, lower share of OPC and higher percentage of HPV positive tumors. Additionally, the follow-up was longer in the training cohort, however a similar percentage of tumor recurrences was observed in both cohorts. Overall, we showed that a radiomic signature might be a useful input in risk assessment in addition to clinical parameters. An additional
RI PT
validation could confirm or contradict the benefit of including radiomics to the modeling. In our study, we observed that heterogeneity of tumor’s CT density is linked more closely to the local tumor control than tumor shape, including tumor volume. Tumors, which are
SC
characterized by more homogenous CT density have a higher probability of local control (Appendix Figure 4S). The HHH large zone high gray-level emphasis was lower in the group
M AN U
of controlled tumors, which corresponds to less patches of large coefficients in the image transformed with 3D high-pass filter. Additionally, the LLL sum entropy was also lower, which relates to a more repeatable pattern of intensity changes in the neighboring voxels. The biology behind this more homogenous texture is unknown but warrants further investigation.
TE D
It can only be speculated that texture heterogeneity might be related to differential biology within the tumor. On the biological level a more heterogeneous microenvironment, e.g. related to tumor perfusion, necrosis or tumor hypoxia, is also known to be associated with a
EP
more resistant phenotype, especially in HNSCC. Our findings agree with the findings of a previously published study, which showed that tumor heterogeneity in CT radiomics was
AC C
related to worse overall survival [6]. Our model shows a higher performance than the overall survival model, which suggest that an additional input, for example the radiomics of lymph node metastasis, might improve the latter. Another set of CT radiomic features was found to be associated with p16 staining. The integration in the host genome and the expression of the viral protein E7, via inactivation of the Retinoblastoma protein, leads to an up-regulation of p16INK4a and correlates strongly with active HPV infection. Therefore, the immunohistochemical detection of the expression level of p16INK4a is nowadays a widely accepted surrogate marker for transcriptionally active HPV
ACCEPTED MANUSCRIPT infections [18]. Our results show that HPV positive tumors seem to be more homogenous in CT density (Appendix Figure 5S). The HPV positive tumors were characterized with significantly lower coefficient of variation and LLL standard deviation, which refers to a smaller width of density distribution in the tumor. They also presented lower LLL small zone
RI PT
high gray-level emphasis, corresponding to lower number of small regions with high intensity. The more homogenous density of HPV-positive tumors also seems to fit to the hypothesis that tumor homogeneity is correlated to a more favorable phenotype of HNSCC.
SC
Our results, observed in two independent patient cohorts, support the hypothesis presented in the explanatory studies [12, 13], that HPV infection is related to tumor density homogeneity.
M AN U
Although both local tumor recurrence and HPV radiomic signatures refer to heterogeneity in the tumor CT density, they consisted of different radiomic features. No strong correlation between features in different models was observed. The local recurrence signature was also prognostic in the subgroup of HPV negative patients (Figure 2S). It suggests that the local
TE D
recurrence radiomic signature is not driven by the HPV status.
Certain limitations were present in this study. The analysis was based on single-center data, although patients enrolled in this study were divided into training and validation cohort as
EP
well as the scanning protocol was not standardized for the training cohort and three different types of CT scanners were used. The proposed models should be further validated in an
AC C
external cohort to test them against different contouring protocols and different scanners. The inter-observer variability in GTV definition in head and neck cancer is a known issue [19], however it influence on CT radiomics remains unclear. Additionally the stability of the features should be evaluated in a test-retest study. Further, patients in training and validation cohorts were treated with different forms of concurrent systemic therapy, either with cisplatin or cetuximab weekly in the training cohort and with both agents, cisplatin and cetuximab, in the consolidation cohort. Of note, the radiomics model seemed to be robust against the choice of concurrent systemic therapy. Moreover, the scanning grid size varied among the patients,
ACCEPTED MANUSCRIPT which is known to influence stability of texture [20]. However, there is no consensus in the literature about the best correction method; in our study the resampling approach to isotropic voxels using linear interpolation was used. These results cannot be directly translated to any other resolution. The investigation of HPV influence on the tumor structure in CT images is
radiomics to histology markers and gene expression are needed.
RI PT
just a one step in understanding tumor heterogeneity in CT. Further studies correlating CT
In conclusion, this study shows that tumor heterogeneity, quantified using contrast-enhanced
SC
CT radiomics, is associated with local tumor control and HPV status in HNSCC. The radiomics-based local tumor control model showed a better performance than models
M AN U
incorporating clinical parameters in our validation cohort. However, this result should be further investigated on a larger, external validation set to provide more evidence on the CT radiomics prognostic power.
TE D
Figure 1. Radiomics-based local tumor control model. The model consists of three radiomic features - HHH large size high gray-level emphasis, LLL sum entropy, and LLH difference
EP
variance. Survival curves split significantly (G-rho test p-value < 0.001) in both training and
AC C
validation cohorts based on the optimal sensitivity-specificity threshold.
Figure 2. The receiver operating characteristics for HPV status prediction using radiomics in training (AUC = 0.85) and validation cohort (AUC = 0.78).
References
ACCEPTED MANUSCRIPT [1]
Ang KK, Harris J, Wheeler R, et al. Human papillomavirus and survival of patients
with oropharyngeal cancer. N Engl J Med 2010;363(1):24-35 [2]
Lassen P, Primdahl H, Johansen J, et al. Impact of HPV-associated p16-expression on
radiotherapy outcome in advanced oropharynx and non-oropharynx cancer. Radiother Oncol
[3]
RI PT
2014;113(3):310-316
Lambin P, Rios-Velazquez E, Leijenaar RTH, et al. Radiomics: extracting more
information from medical images using advanced feature analysis. Eur J Cancer
[4]
SC
2012;48(4):441-446
Yip SS, Aerts HJ. Applications and limitations of radiomics. Phys Med Biol
[5]
M AN U
2016;61(13):150-166
Hatt M, Tixier F, Pierce L, Kinahan PE, Le Rest CC, Visvikis D. Characterization of
PET/CT images using texture analysis: the past, the present… any future? Eur J Nucl Med Mol Imaging 2016
Aerts HJWL, Velazquez ER, Leijenaar RTH, et al. Decoding tumour phenotype by
noninvasive
imaging
2014;5(4006):4006
using
a
quantitative
radiomics
approach.
Nat
Commun
Yang D, Rao G, Martinez J, Veeraraghavan A, Rao A. Evaluation of tumor-derived
EP
[7]
TE D
[6]
MRI-texture features for discrimination of molecular subtypes and prediction of 12-month
[8]
AC C
survival status in glioblastoma. Med Phys 2015;42(11):6725-6735 Wong AJ, Kanwar A, Mohamed AS, Fuller CD. Radiomics in head and neck cancer:
from exploration to application. Translational Cancer Research 2016;5(4) [9]
Leijenaar RTH, Carvalho S, Hoebers FJ, et al. External validation of a prognostic CT-
based radiomic signature in oropharyngeal squamous cell carcinoma. Acta Oncol 2015;54(9): 1423-1429
ACCEPTED MANUSCRIPT [10]
Parmar C, Leijenaar RTH, Grossmann P, et al. Radiomic feature clusters and
prognostic signatures
specific for
Lung
and
Head
&
Neck
cancer.
Sci
Rep
2015;5(11004):11004 [11]
Harrison LB, Sessions RB, Hong WK. Head and neck cancer: a multidisciplinary
[12]
RI PT
approach. Lippincott Williams & Wilkins, Wolters Kluwer, Philadelphia, 2009.
Buch K, Fujita A, Li B. Using Texture Analysis to Determine Human Papillomavirus
Status of Oropharyngeal Squamous Cell Carcinomas on CT. AJNR Am J Neuroradiol
[13]
SC
2015;36:1343-1348
Fujita A, Buch K, Li B, Kawashima Y, Qureshi MM, Sakai O. Difference Between
M AN U
HPV-Positive and HPV-Negative Non-Oropharyngeal Head and Neck Cancer: Texture Analysis Features on CT. J Comput Assist Tomogr 2016;40(1) [14]
Horn JL. A rationale and test for the number of factors in factor analysis.
Psychometrika 1965;30:179-185
R. D. C. Team, R: A Language and Environment for Statistical Computing, Vienna,
TE D
[15]
Austria: R Foundation for Statistical Computing, 2010. [16]
F. E. Harrell. Regression Modeling Strategies: With Applications to Linear Models,
EP
Logistic Regression, and Survival Analysis, Springer, 2011. [17] Michiels S, Le Maître A, Buyse M, et al. Surrogate endpoints for overall survival in
AC C
locally advanced head and neck cancer: meta-analyses of individual patient data. Lancet Oncol. 2009;10:341-350 [18]
Krupar R, Hartl M, Wirsching K, Dietmaier W, Strutz J, Hofstaedter F. Comparison of
HPV prevalence in HNSCC patients with regard to regional and socioeconomic factors. Eur Arch Otorhinolaryngol 2014;271:1737-1745 [19] Riegel AC, Berson AM, Destian S, et al. Variability of gross tumor volume delineation in head-and-neck cancer using CT and PET/CT fusion. Int J Radiat Oncol Biol Phys. 2006;65(3):726-732
ACCEPTED MANUSCRIPT [20] Zhao B, Tan Y, Tsai WY, Qi J, Xie C, Lu L, Schwartz LH. Reproducibility of radiomics
AC C
EP
TE D
M AN U
SC
RI PT
for deciphering tumor phenotype with imaging. Sci Rep 2016;6(23428):23428
ACCEPTED MANUSCRIPT Number of patients
Training cohort
Validation cohort
93
56
Number of recurrences 24 (25%)
14 (25%)
47 (3-156)
16 (3-28)
p16 positive
36 (39%)
26 (46%)
p16 negative
57 (61%)
Median follow-up
33 (36%)
T3/T4
60 (64%)
N0
15 (16%)
N1
5 (5%)
N2
69 (74%)
N3
4 (5%)
Nodal stage
Tumor site
Oropharynx Hypopharynx Larynx Oral cavity Female
4 (7%) 40 (71%)
2 (4%)
70 (75%)
30 (54%)
18 (19%)
8 (14%)
4 (5%)
9 (16%)
1 (1%)
9 (16%)
21 (23%)
13 (23%) 43 (77%)
Age [years]*
61.9 (41.9 – 74.5)
59.8 (47.2 – 75.7)
CT scanners
Siemens SOMATOM
Siemens SOMATOM
Volume Zoom (n=64)
Definition AS (n=56)
EP
AC C
parameters
10 (18%)
72 (77%)
Male
Scanning
47 (84%)
TE D
Gender
9 (16%)
M AN U
Tumor stage T1/T2
30 (54%)
SC
HPV status
RI PT
[months]
Siemens SOMATOM Definition AS (n=22) GE Discovery STE (n=5)
Slice thickness [mm]
1.25-3.30
2
In-plane resolution
0.85-1.95
0.98-1.56
kV
120; 140
120
mAs
60-450
183-450
[mm]
Table 1. Studied patient cohorts and imaging protocol details. * median and the range in the parentheses.
ACCEPTED MANUSCRIPT Feature group Feature name
CI
FDR
1
LLL sum entropy
0.75
0.11
2
LHL IMC1
0.67
0.12
3
HHH large zone high 0.68
0.13
0.19
difference 0.65
0.16
HHL contrastNGTDM
5
LLH variance
HLH complexity
0.64
0.16
7
Sum entropy
0.64
0.16
8
LHH MCC
0.68
0.12
9
LHL low gray-level 0.70
0.12
emphasis
M AN U
6
SC
0.69
4
RI PT
gray-level emphasis
Table 2. List of uncorrelated (independent) radiomic features prognostic for local tumor control in the univariate Cox regression analysis, CI – concordance index, FDR – false discovery rate. The features selected in the multivariate Cox regression analysis are
AC C
EP
TE D
underlined.
ACCEPTED MANUSCRIPT Training cohort Model
Parameters of the model
Radiomics
HHH
large
zone
CI
high
gray-level 0.75
emphasis
Validation cohort
p-value CI
p-value
< 0.001
< 0.001
(0.74-0.76)
0.78 (0.77-0.79)
LLL sum entropy
Clinical
stage
0.79
tumor volume
(0.78-0.80)
HPV HHH
large
zone
high
gray-level 0.75
emphasis
(0.74-0.76)
Radiomics
LLH difference variance
+ clinical
tumor volume HPV
M AN U
LLL sum entropy LLH difference variance
0.73
(0.71-0.75)
< 0.001
SC
Radiomics
RI PT
LLH difference variance
0.78
< 0.046
(0.77-0.79)
0.80
0.76
(0.79-0.81)
(0.75-0.77)
Table 3. Comparison of the prognostic models for local tumor control, CI - concordance index
TE D
and its 95% confidence interval in brackets; Wilcoxon test p-values are given for the
AC C
EP
comparison of the models performance.
ACCEPTED MANUSCRIPT Feature
Feature name
AUC
FDR
1
LLL gray-level non-uniformity
0.77
0.053
2
LLL standard deviation
0.83
0.002
3
difference entropy
0.76
0.004
4
LLL small zone high gray-level 0.78
0.004
5
HHL difference entropy
0.67
6
compactness 2
0.66
7
standard deviation
0.79
8
coefficient of variation
9
spherical disproportion
SC
emphasis
RI PT
group
0.143
0.077
0.004
0.69
0.071
0.66
0.069
M AN U
Table 4. List of uncorrelated (independent) radiomic features predictive for HPV status. The features selected in the multivariate logistic regression are underlined. FDR – false discovery
AC C
EP
TE D
rate.
RI PT M AN U
SC
0.6 0.4 0
5
AC C
EP
TE D
0.2
< threshold validation >= threshold validation < threshold training >= threshold training
0.0
Local control rate
0.8
1.0
ACCEPTED MANUSCRIPT
10
15
20
25
Months
59
57
54
47
46
42
< threshold training
34
27
20
15
12
11
>= threshold training
37
34
27
22
17
3
< threshold validation
19
13
9
9
8
1
>= threshold validation
M AN U
0.6
TE D
0.4
EP
0.2
AC C
0.0
Sensitivity
0.8
SC
1.0
RI PT
ACCEPTED MANUSCRIPT
1.0
0.8
training validation
0.6
0.4
Specificity
0.2
0.0
S0360-3016(17)31011-8
DOI:
10.1016/j.ijrobp.2017.06.002
Reference:
ROB 24297
To appear in:
International Journal of Radiation Oncology • Biology • Physics
Received Date: 22 December 2016 Revised Date:
15 May 2017
Accepted Date: 5 June 2017
Please cite this article as: Bogowicz M, Riesterer O, Ikenberg K, Stieb S, Moch H, Studer G, Guckenberger M, Tanadini-Lang S, CT radiomics predicts HPV status and local tumor control after definitive radiochemotherapy in head and neck squamous cell carcinoma, International Journal of Radiation Oncology • Biology • Physics (2017), doi: 10.1016/j.ijrobp.2017.06.002. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
CT radiomics predicts HPV status and local tumor control after definitive radiochemotherapy in head and neck squamous cell carcinoma
Marta Bogowicz, MSc, 1Oliver Riesterer, MD, 2Kristian Ikenberg, MD, 1Sonja Stieb, MD, 2Holger
RI PT
1
Moch, MD, Prof, 1Gabriela Studer, MD, Prof, 1Matthias Guckenberger, MD, Prof, 1Stephanie TanadiniLang, PhD
Department of Radiation Oncology and 2Department of Pathology and Molecular Pathology, University
Hospital Zurich, University of Zurich, Switzerland
Address: Department of Radiation Oncology University Hospital Zurich Rämistrasse 100
Switzerland
Corresponding author: Marta Bogowicz
TE D
8091 Zurich
M AN U
Running title: Radiomics predicts HPV and tumor control
SC
1
EP
Email: [email protected] Phone: +41 44 255 4116
AC C
Fax: +41 44 255 4547
Conflict of interest: Authors declare no conflict of interest.
Acknowledgments:
The project was supported by the Clinical Research Priority Program Tumor Oxygenation of the University of Zurich, by a grant of the Matching Fund of the University of Zurich and by a research grant from Merck (Schweiz) AG.
ACCEPTED MANUSCRIPT
Summary: This study shows that higher pre-treatment tumor heterogeneity, quantified using radiomics of contrast-enhanced CT imaging, is associated with worse prognosis in head and neck squamous
RI PT
cell carcinoma and the absence of HPV infection. Overall, our analysis provides an evidence that a radiomic signature might be a useful input in risk assessment in addition to clinical parameters
AC C
EP
TE D
M AN U
SC
(tumor stage, tumor volume and HPV status).
ACCEPTED MANUSCRIPT CT radiomics predicts HPV status and local tumor control after definitive radiochemotherapy in head and neck squamous cell carcinoma
RI PT
Running title: Radiomics predicts HPV and tumor control
Abstract Purpose
SC
This study aimed to predict local tumor control (LC) after radiochemotherapy of head and neck squamous cell carcinoma (HNSCC) and HPV status using CT radiomics.
M AN U
Materials and methods
HNSCC patients treated with definitive radiochemotherapy were included in the retrospective study approved by local ethical commission (93 and 56 patients in the training and validation cohorts, respectively). 317 CT radiomic features, including shape-, intensity-, texture- and transform-based features, were calculated in the primary tumor region. A Cox and logistic
TE D
regression models were built to predict LC and HPV status, respectively. The best performing features in the univariate analysis were included in the multivariate analysis, after exclusion
EP
of redundant features. The quality of the models was assessed using the concordance index (CI) for modelling of LC and receiver operating characteristics area under the curve (AUC)
AC C
for HPV status prediction. The radiomics LC model was compared to model incorporating clinical parameters (tumor stage, volume and HPV status) and a mixed model. Results
A radiomic signature, comprising three features, was significantly associated with LC (CItraining=0.75 and CIvalidation=0.78), showing that tumors with a more heterogeneous CT density distribution are at risk for decreased LC. Addition of clinical parameters to the radiomics model slightly improved the model in the training but not in validation cohort. Another radiomic signature showed a good performance in HPV status prediction
ACCEPTED MANUSCRIPT (AUCtraining=0.85 and AUCvalidation=0.78) and indicated that HPV positive tumors have more homogenous CT density distribution. Conclusion HNSCC tumor density heterogeneity, quantified by CT radiomics, is associated with LC after
RI PT
radiochemotherapy and HPV status.
SC
Introduction
Head and neck squamous cell carcinoma (HNSCC) is the fifth most common cancer
M AN U
worldwide and the standard of care for advanced unresectable cases is definitive radiochemotherapy. Clinical outcome after radiochemotherapy is very heterogeneous ranging from about 80% locoregional control and 5 years overall survival in HPV-positive oropharyngeal carcinoma (OPC) to below 50% in HPV-negative OPC and non-oropharyngeal
TE D
cancers [1, 2]. In all cases distant metastasis is infrequent and mainly a consequence of locoregional treatment failure. Therefore, improvement of radiochemotherapy is essential. An important approach is the development of tumor biomarkers that can predict locoregional
EP
tumor control in order to identify patients for treatment intensification. Radiomics is a promising tool for non-invasive characterization of tumor phenotype. It
AC C
extracts a large number of quantitative features from radiographic images: regions of interest are characterized regarding their shape, intensities and texture of both spatial and frequency domains (for example using a wavelet transform) [3]. A correlation between radiomic features and treatment response has been shown using different image modalities and for different tumor types [4, 5]. To provide a biological interpretation of radiomics several groups tried to relate a radiomic signature (set of radiomic parameters) to differential gene expressions including tumor proliferation and molecular subtypes [6, 7].
ACCEPTED MANUSCRIPT Although several CT radiomics-based prognostic models have been proposed for HNSCC [8], it is not fully understood why a particular radiomic phenotype is associated with treatment response. CT radiomics of the primary tumor was successfully used to stratify patients into low and high risk groups in terms of patient overall survival [6, 9, 10, 8]. It was shown to
RI PT
further improve survival models based on clinical parameters, such as TNM stage and tumor volume [6]. While overall survival is an important clinical endpoint, it is also influenced by non-tumor related factors [11], which cannot be captured by pretreatment radiomics of the
SC
primary tumor. However, especially in the case of HNSCC, it is important to understand the influence of intra-tumor heterogeneity on the local response to treatment. Therefore, a study
M AN U
using local tumor control as endpoint is of high clinical interest. Furthermore, the correlation of radiomic phenotype to the important outcome prognostic factor HPV is of interest and links image analysis with tumor biology [12, 13]. Exploratory studies have already indicated a relationship between HPV infection and the heterogeneity of imaging-based tumor density
TE D
[12, 13]. However, these studies were based on small patient numbers and lacked validation cohorts.
In this study we investigated, if CT-based radiomic phenotype of HNSCC is correlated with
EP
local tumor control after radiochemotherapy. Furthermore, using two big patient cohorts we
AC C
studied the influence of HPV infection on tumor radiomics.
Materials and methods
Studied population
This study was a retrospective analysis approved by a local ethical commission. The training cohort consisted of 93 stage III and IV HNSCC patients treated between 2003-2013 with definitive radiotherapy (on average 70 Gy) either in combination with cisplatin (40 mg/m2, up
ACCEPTED MANUSCRIPT to 7 cycles) or cetuximab (loading dose 400 mg/m2 followed by 250 mg/m2 weekly). Induction chemotherapy and/or prior surgery (biopsy allowed) were exclusion criteria. Additionally, data from 56 patients with the same tumor stages from an institutional phase III prospective study (NCT01435252) with a standardized planning CT imaging protocol were
RI PT
selected as validation cohort. In the latter study, patients were treated with triple therapy consisting of radiotherapy and weekly concurrent cisplatin/cetuximab (same doses as in training cohort) with or without consolidation cetuximab (500 mg/m2 biweekly x 6).
SC
Radiotherapy was delivered using intensity modulated technique in all patients.
All patients underwent pretreatment contrast-enhanced planning CT imaging. Patient
M AN U
characteristics and imaging protocols are presented in the Table 1. In all patients the HPV status was determined by immunohistochemical p16 staining. Local recurrence was defined as biopsy proven recurrence within the high dose PTV of the primary tumor.
Follow-up
TE D
consisted of 2-3 monthly examinations by ENT specialists and repetitive FDG-PET/CTs.
Radiomics analysis and image post-processing
EP
An in-house developed radiomics software implementation written in Python programing language was used. It provides 3D analysis of medical images using all four feature extraction
• • •
AC C
methods:
shape (n = 11)
intensity (n = 5)
texture: the Gray Level Co-occurrence Matrix (n = 14), the Neighborhood Gray Tone Difference Matrix (n = 4) and the Gray Level Size Zone Matrix (n = 11)
•
wavelet transform using Coiflet function (8 sub-bands, n = 272).
In total 317 features were extracted per patient. For intensity, texture and wavelet analysis images were resized to the symmetrical voxels of size of 3.3 mm, corresponding to the lowest
ACCEPTED MANUSCRIPT image resolution (sagittal) in the training cohort. Additionally, Hounsfield Units range of -20 to 180 HU was applied to limit the analysis to soft tissue only. To quantify the texture images were resampled to equally spaced bins of 5 HU. A detailed description of radiomics implementation is presented in the Appendix.
RI PT
The tumor heterogeneity was studied inside the primary GTV defined on the contrastenhanced planning CT by a radiation oncologist with more than 10 years of experience. Contours were post-processed to account for metal artifacts occurring in the CT images.
SC
Contours were removed from artifact-affected slices. Scans with more than a half of the
Radiomic inter-features correlations
M AN U
contoured slices affected by metal artifacts were not included in the analysis.
Radiomic features were grouped according to principal component analysis using data from
TE D
the training cohort to account for the inter-features correlations. The Horn method was used to choose the optimal number of retained components [14]. The contribution of radiomic features to principal components were determined. Features were assigned to a principal
EP
component to which they contributed the most. Each principal component corresponded to
AC C
one group of correlated features. The statistical analysis was performed in R (v. 3.2.4) [15].
Radiomics prognostic value of local tumor control
Radiomic features prognostic for local tumor control were preselected in the univariate Cox regression analysis. The results were corrected for multiple testing with false discovery rate (FDR) < 0.20. Additionally, the correlated features were excluded from the further analysis by selecting only the most prognostic parameter per principal component feature group. Feature prognostic value was quantified using concordance index (CI) [16]. The preselected features
ACCEPTED MANUSCRIPT were entered in the multivariate Cox regression analysis. The radiomics-based model was compared with a multivariate model created with clinical parameters: tumor volume, T stage (T1/T2 vs T3/T4) and HPV status. Finally, a mixed model was created, where both radiomic features and clinical parameters were incorporated in the multivariate Cox model. All
RI PT
multivariate models were built using Akaike information criterion in the backward selection of variables. The prognostic value of obtained models was compared based on the CI, using Wilcoxon test (p-value < 0.05) and the bootstrap method with 100 randomly selected samples
SC
to calculate the CI distribution. The results were verified in the validation cohort. Curves were split based on the optimal threshold from the receiver operating characteristic curve at 18
M AN U
months maximizing both sensitivity and specificity and compared using G-rho test (pvalue < 0.05).
TE D
Prediction of HPV status
The radiomic features for HPV status prediction were also preselected using principal component analysis in combination with univariate logistic regression. The univariate model
EP
accuracy was quantified by the area under receiver operating characteristic curve (AUC). The results were corrected for multiple testing with FDR < 0.20. A radiomic feature characterized
AC C
by the highest AUC was selected in each principal component group. This set was later used in multivariate logistic regression with backward selection of the variables. The cutoff threshold of the final model was set to maximize both sensitivity and specificity in the training cohort. The results were again verified in the validation cohort.
ACCEPTED MANUSCRIPT
Results
RI PT
Radiomics-based prognostic model for local tumor control
Nine components were selected in the principal component analysis, which lead to nine groups of correlated radiomic features. The group sizes varied from 21 to 66 features. In the
SC
univariate Cox regression analysis 83 out of 317 radiomic features were found to be prognostic for local tumor control (FDR < 0.20). The list of the most prognostic features,
M AN U
accounting for inter-feature correlations, is shown in Table 2. In the multivariate analysis with the backward selection of variables three (HHH large zone high gray-level emphasis, LLL sum entropy, and LLH difference variance) out of nine previously mentioned features were selected, resulting in a model with CI of 0.75. This model was found to be prognostic also in
TE D
the validation cohort with CI = 0.78 and significant for stratification into groups with low and high risk of recurrence (Figure 1; Table 3). Additionally, the model was prognostic in the subgroup analysis of oropharyngeal carcinoma as well as HPV negative tumors (Appendix
EP
Table 1S and Figure 2S). Different CT scanners and kV settings did not show a major
AC C
influence on the radiomic features included in the model (Appendix Figure 6S).
Comparison of the radiomics-based and clinical-based prognostic model for local tumor control
The radiomics-based model was compared with a model incorporating known prognostic clinical parameters (tumor volume, T stage and HPV status) as well as with a mixed-model (radiomics and clinical parameters). The parameters chosen in the multivariate analysis and the results of Wilcoxon test, studying differences between the models, are presented in the
ACCEPTED MANUSCRIPT Table 3. The combination of radiomics (LLH difference variance) and clinical parameters (HPV and tumor volume) was found to have the highest prognostic power (CI = 0.80) in the training cohort, however performed worse than the radiomics-based model in the validation cohort. The same trend was observed for the subgroup of oropharyngeal carcinoma (Appendix
RI PT
Table 1S). In the mixed model the single radiomic feature was not significant however, it had a significant impact on the improvement of the model discriminative power (Table 2S).
SC
Prediction of HPV status
M AN U
In total, 86 radiomic features were found to be predictive for HPV status (FDR < 0.20). Per group of correlated features, the one characterized with the highest AUC was chosen (Table 4) and entered in the multivariate logistic regression. The final model comprised radiomic features: LLL standard deviation
•
LLL small zone high gray-level emphasis
•
HHL difference entropy
•
coefficient of variation
EP
TE D
•
four
AC C
The final radiomic model for HPV status prediction showed a good performance in both training (AUC = 0.85) and validation cohort (AUC = 0.78). The specificity of the model was higher in the validation in comparison to the training cohort (specificitytraining = 0.72 and specificityvalidation = 0.82), whereas the sensitivity was lower (sensitivitytraining = 0.72 and sensitivitytraining = 0.62). Figure 2 presents the receiver operating characteristics for the model. The HPV radiomic signature was independent from the local control signature as radiomic features in these two models were not correlated and the HPV prediction model was valid in patient subgroups regardless of the tumor control (Appendix Figure 3S). Similar to the local
ACCEPTED MANUSCRIPT tumor control model, we have not observed major variations in our HPV radiomic signature depending on the CT scanner and kV settings (Appendix Figure 7S). Discussion This study evaluated the prognostic value of contrast enhanced CT-based radiomics in head
RI PT
and neck squamous cell carcinoma regarding local tumor control and HPV status. The radiomics-based local control model was compared to a model based on prognostic clinical parameters (tumor stage, tumor volume and HPV status) and to the mixed model (clinical plus
SC
radiomics). The performance of the models was verified in an independent cohort of patients. Although the radiomics-based model achieved lower concordance index in comparison to the
M AN U
mixed model in the training cohort, it showed a significantly higher performance in the validation cohort. The stable performance of radiomics-based model in both training and validation patient cohorts stresses its high credibility. Radiomics was also shown to be closely associated with the HPV status with AUC values of 0.85 and 0.78 in the training and the
TE D
validation cohort, respectively.
In the case of HNSCC, locoregional control closely correlates with overall survival [17]. Prediction of local treatment efficacy is of high interest in order to modify treatment
EP
according to the individual tumor’s radiosensitivity. The radiomics model created by us was highly prognostic for local tumor control and addition of the clinically relevant biomarker,
AC C
HPV status, could only slightly improve the model in the training cohort but not in the validation cohort. Similar results were observed for the analysis of OPC patients only (Appendix Table 1S), which indicates slight overfitting of the mixed model (radiomics and clinical predictors) in the training cohort. The value of the radiomic feature (LLH difference variance) in the mixed model should be studied in more details. This covariate was not significant in the multivariate modeling however, had a positive impact on the discriminative power of the model, especially in the validation cohort. The training and validation cohorts were not perfectly matched. The validation cohort was characterized by higher percentage of
ACCEPTED MANUSCRIPT more advanced tumors, lower share of OPC and higher percentage of HPV positive tumors. Additionally, the follow-up was longer in the training cohort, however a similar percentage of tumor recurrences was observed in both cohorts. Overall, we showed that a radiomic signature might be a useful input in risk assessment in addition to clinical parameters. An additional
RI PT
validation could confirm or contradict the benefit of including radiomics to the modeling. In our study, we observed that heterogeneity of tumor’s CT density is linked more closely to the local tumor control than tumor shape, including tumor volume. Tumors, which are
SC
characterized by more homogenous CT density have a higher probability of local control (Appendix Figure 4S). The HHH large zone high gray-level emphasis was lower in the group
M AN U
of controlled tumors, which corresponds to less patches of large coefficients in the image transformed with 3D high-pass filter. Additionally, the LLL sum entropy was also lower, which relates to a more repeatable pattern of intensity changes in the neighboring voxels. The biology behind this more homogenous texture is unknown but warrants further investigation.
TE D
It can only be speculated that texture heterogeneity might be related to differential biology within the tumor. On the biological level a more heterogeneous microenvironment, e.g. related to tumor perfusion, necrosis or tumor hypoxia, is also known to be associated with a
EP
more resistant phenotype, especially in HNSCC. Our findings agree with the findings of a previously published study, which showed that tumor heterogeneity in CT radiomics was
AC C
related to worse overall survival [6]. Our model shows a higher performance than the overall survival model, which suggest that an additional input, for example the radiomics of lymph node metastasis, might improve the latter. Another set of CT radiomic features was found to be associated with p16 staining. The integration in the host genome and the expression of the viral protein E7, via inactivation of the Retinoblastoma protein, leads to an up-regulation of p16INK4a and correlates strongly with active HPV infection. Therefore, the immunohistochemical detection of the expression level of p16INK4a is nowadays a widely accepted surrogate marker for transcriptionally active HPV
ACCEPTED MANUSCRIPT infections [18]. Our results show that HPV positive tumors seem to be more homogenous in CT density (Appendix Figure 5S). The HPV positive tumors were characterized with significantly lower coefficient of variation and LLL standard deviation, which refers to a smaller width of density distribution in the tumor. They also presented lower LLL small zone
RI PT
high gray-level emphasis, corresponding to lower number of small regions with high intensity. The more homogenous density of HPV-positive tumors also seems to fit to the hypothesis that tumor homogeneity is correlated to a more favorable phenotype of HNSCC.
SC
Our results, observed in two independent patient cohorts, support the hypothesis presented in the explanatory studies [12, 13], that HPV infection is related to tumor density homogeneity.
M AN U
Although both local tumor recurrence and HPV radiomic signatures refer to heterogeneity in the tumor CT density, they consisted of different radiomic features. No strong correlation between features in different models was observed. The local recurrence signature was also prognostic in the subgroup of HPV negative patients (Figure 2S). It suggests that the local
TE D
recurrence radiomic signature is not driven by the HPV status.
Certain limitations were present in this study. The analysis was based on single-center data, although patients enrolled in this study were divided into training and validation cohort as
EP
well as the scanning protocol was not standardized for the training cohort and three different types of CT scanners were used. The proposed models should be further validated in an
AC C
external cohort to test them against different contouring protocols and different scanners. The inter-observer variability in GTV definition in head and neck cancer is a known issue [19], however it influence on CT radiomics remains unclear. Additionally the stability of the features should be evaluated in a test-retest study. Further, patients in training and validation cohorts were treated with different forms of concurrent systemic therapy, either with cisplatin or cetuximab weekly in the training cohort and with both agents, cisplatin and cetuximab, in the consolidation cohort. Of note, the radiomics model seemed to be robust against the choice of concurrent systemic therapy. Moreover, the scanning grid size varied among the patients,
ACCEPTED MANUSCRIPT which is known to influence stability of texture [20]. However, there is no consensus in the literature about the best correction method; in our study the resampling approach to isotropic voxels using linear interpolation was used. These results cannot be directly translated to any other resolution. The investigation of HPV influence on the tumor structure in CT images is
radiomics to histology markers and gene expression are needed.
RI PT
just a one step in understanding tumor heterogeneity in CT. Further studies correlating CT
In conclusion, this study shows that tumor heterogeneity, quantified using contrast-enhanced
SC
CT radiomics, is associated with local tumor control and HPV status in HNSCC. The radiomics-based local tumor control model showed a better performance than models
M AN U
incorporating clinical parameters in our validation cohort. However, this result should be further investigated on a larger, external validation set to provide more evidence on the CT radiomics prognostic power.
TE D
Figure 1. Radiomics-based local tumor control model. The model consists of three radiomic features - HHH large size high gray-level emphasis, LLL sum entropy, and LLH difference
EP
variance. Survival curves split significantly (G-rho test p-value < 0.001) in both training and
AC C
validation cohorts based on the optimal sensitivity-specificity threshold.
Figure 2. The receiver operating characteristics for HPV status prediction using radiomics in training (AUC = 0.85) and validation cohort (AUC = 0.78).
References
ACCEPTED MANUSCRIPT [1]
Ang KK, Harris J, Wheeler R, et al. Human papillomavirus and survival of patients
with oropharyngeal cancer. N Engl J Med 2010;363(1):24-35 [2]
Lassen P, Primdahl H, Johansen J, et al. Impact of HPV-associated p16-expression on
radiotherapy outcome in advanced oropharynx and non-oropharynx cancer. Radiother Oncol
[3]
RI PT
2014;113(3):310-316
Lambin P, Rios-Velazquez E, Leijenaar RTH, et al. Radiomics: extracting more
information from medical images using advanced feature analysis. Eur J Cancer
[4]
SC
2012;48(4):441-446
Yip SS, Aerts HJ. Applications and limitations of radiomics. Phys Med Biol
[5]
M AN U
2016;61(13):150-166
Hatt M, Tixier F, Pierce L, Kinahan PE, Le Rest CC, Visvikis D. Characterization of
PET/CT images using texture analysis: the past, the present… any future? Eur J Nucl Med Mol Imaging 2016
Aerts HJWL, Velazquez ER, Leijenaar RTH, et al. Decoding tumour phenotype by
noninvasive
imaging
2014;5(4006):4006
using
a
quantitative
radiomics
approach.
Nat
Commun
Yang D, Rao G, Martinez J, Veeraraghavan A, Rao A. Evaluation of tumor-derived
EP
[7]
TE D
[6]
MRI-texture features for discrimination of molecular subtypes and prediction of 12-month
[8]
AC C
survival status in glioblastoma. Med Phys 2015;42(11):6725-6735 Wong AJ, Kanwar A, Mohamed AS, Fuller CD. Radiomics in head and neck cancer:
from exploration to application. Translational Cancer Research 2016;5(4) [9]
Leijenaar RTH, Carvalho S, Hoebers FJ, et al. External validation of a prognostic CT-
based radiomic signature in oropharyngeal squamous cell carcinoma. Acta Oncol 2015;54(9): 1423-1429
ACCEPTED MANUSCRIPT [10]
Parmar C, Leijenaar RTH, Grossmann P, et al. Radiomic feature clusters and
prognostic signatures
specific for
Lung
and
Head
&
Neck
cancer.
Sci
Rep
2015;5(11004):11004 [11]
Harrison LB, Sessions RB, Hong WK. Head and neck cancer: a multidisciplinary
[12]
RI PT
approach. Lippincott Williams & Wilkins, Wolters Kluwer, Philadelphia, 2009.
Buch K, Fujita A, Li B. Using Texture Analysis to Determine Human Papillomavirus
Status of Oropharyngeal Squamous Cell Carcinomas on CT. AJNR Am J Neuroradiol
[13]
SC
2015;36:1343-1348
Fujita A, Buch K, Li B, Kawashima Y, Qureshi MM, Sakai O. Difference Between
M AN U
HPV-Positive and HPV-Negative Non-Oropharyngeal Head and Neck Cancer: Texture Analysis Features on CT. J Comput Assist Tomogr 2016;40(1) [14]
Horn JL. A rationale and test for the number of factors in factor analysis.
Psychometrika 1965;30:179-185
R. D. C. Team, R: A Language and Environment for Statistical Computing, Vienna,
TE D
[15]
Austria: R Foundation for Statistical Computing, 2010. [16]
F. E. Harrell. Regression Modeling Strategies: With Applications to Linear Models,
EP
Logistic Regression, and Survival Analysis, Springer, 2011. [17] Michiels S, Le Maître A, Buyse M, et al. Surrogate endpoints for overall survival in
AC C
locally advanced head and neck cancer: meta-analyses of individual patient data. Lancet Oncol. 2009;10:341-350 [18]
Krupar R, Hartl M, Wirsching K, Dietmaier W, Strutz J, Hofstaedter F. Comparison of
HPV prevalence in HNSCC patients with regard to regional and socioeconomic factors. Eur Arch Otorhinolaryngol 2014;271:1737-1745 [19] Riegel AC, Berson AM, Destian S, et al. Variability of gross tumor volume delineation in head-and-neck cancer using CT and PET/CT fusion. Int J Radiat Oncol Biol Phys. 2006;65(3):726-732
ACCEPTED MANUSCRIPT [20] Zhao B, Tan Y, Tsai WY, Qi J, Xie C, Lu L, Schwartz LH. Reproducibility of radiomics
AC C
EP
TE D
M AN U
SC
RI PT
for deciphering tumor phenotype with imaging. Sci Rep 2016;6(23428):23428
ACCEPTED MANUSCRIPT Number of patients
Training cohort
Validation cohort
93
56
Number of recurrences 24 (25%)
14 (25%)
47 (3-156)
16 (3-28)
p16 positive
36 (39%)
26 (46%)
p16 negative
57 (61%)
Median follow-up
33 (36%)
T3/T4
60 (64%)
N0
15 (16%)
N1
5 (5%)
N2
69 (74%)
N3
4 (5%)
Nodal stage
Tumor site
Oropharynx Hypopharynx Larynx Oral cavity Female
4 (7%) 40 (71%)
2 (4%)
70 (75%)
30 (54%)
18 (19%)
8 (14%)
4 (5%)
9 (16%)
1 (1%)
9 (16%)
21 (23%)
13 (23%) 43 (77%)
Age [years]*
61.9 (41.9 – 74.5)
59.8 (47.2 – 75.7)
CT scanners
Siemens SOMATOM
Siemens SOMATOM
Volume Zoom (n=64)
Definition AS (n=56)
EP
AC C
parameters
10 (18%)
72 (77%)
Male
Scanning
47 (84%)
TE D
Gender
9 (16%)
M AN U
Tumor stage T1/T2
30 (54%)
SC
HPV status
RI PT
[months]
Siemens SOMATOM Definition AS (n=22) GE Discovery STE (n=5)
Slice thickness [mm]
1.25-3.30
2
In-plane resolution
0.85-1.95
0.98-1.56
kV
120; 140
120
mAs
60-450
183-450
[mm]
Table 1. Studied patient cohorts and imaging protocol details. * median and the range in the parentheses.
ACCEPTED MANUSCRIPT Feature group Feature name
CI
FDR
1
LLL sum entropy
0.75
0.11
2
LHL IMC1
0.67
0.12
3
HHH large zone high 0.68
0.13
0.19
difference 0.65
0.16
HHL contrastNGTDM
5
LLH variance
HLH complexity
0.64
0.16
7
Sum entropy
0.64
0.16
8
LHH MCC
0.68
0.12
9
LHL low gray-level 0.70
0.12
emphasis
M AN U
6
SC
0.69
4
RI PT
gray-level emphasis
Table 2. List of uncorrelated (independent) radiomic features prognostic for local tumor control in the univariate Cox regression analysis, CI – concordance index, FDR – false discovery rate. The features selected in the multivariate Cox regression analysis are
AC C
EP
TE D
underlined.
ACCEPTED MANUSCRIPT Training cohort Model
Parameters of the model
Radiomics
HHH
large
zone
CI
high
gray-level 0.75
emphasis
Validation cohort
p-value CI
p-value
< 0.001
< 0.001
(0.74-0.76)
0.78 (0.77-0.79)
LLL sum entropy
Clinical
stage
0.79
tumor volume
(0.78-0.80)
HPV HHH
large
zone
high
gray-level 0.75
emphasis
(0.74-0.76)
Radiomics
LLH difference variance
+ clinical
tumor volume HPV
M AN U
LLL sum entropy LLH difference variance
0.73
(0.71-0.75)
< 0.001
SC
Radiomics
RI PT
LLH difference variance
0.78
< 0.046
(0.77-0.79)
0.80
0.76
(0.79-0.81)
(0.75-0.77)
Table 3. Comparison of the prognostic models for local tumor control, CI - concordance index
TE D
and its 95% confidence interval in brackets; Wilcoxon test p-values are given for the
AC C
EP
comparison of the models performance.
ACCEPTED MANUSCRIPT Feature
Feature name
AUC
FDR
1
LLL gray-level non-uniformity
0.77
0.053
2
LLL standard deviation
0.83
0.002
3
difference entropy
0.76
0.004
4
LLL small zone high gray-level 0.78
0.004
5
HHL difference entropy
0.67
6
compactness 2
0.66
7
standard deviation
0.79
8
coefficient of variation
9
spherical disproportion
SC
emphasis
RI PT
group
0.143
0.077
0.004
0.69
0.071
0.66
0.069
M AN U
Table 4. List of uncorrelated (independent) radiomic features predictive for HPV status. The features selected in the multivariate logistic regression are underlined. FDR – false discovery
AC C
EP
TE D
rate.
RI PT M AN U
SC
0.6 0.4 0
5
AC C
EP
TE D
0.2
< threshold validation >= threshold validation < threshold training >= threshold training
0.0
Local control rate
0.8
1.0
ACCEPTED MANUSCRIPT
10
15
20
25
Months
59
57
54
47
46
42
< threshold training
34
27
20
15
12
11
>= threshold training
37
34
27
22
17
3
< threshold validation
19
13
9
9
8
1
>= threshold validation
M AN U
0.6
TE D
0.4
EP
0.2
AC C
0.0
Sensitivity
0.8
SC
1.0
RI PT
ACCEPTED MANUSCRIPT
1.0
0.8
training validation
0.6
0.4
Specificity
0.2
0.0