Header

UZH-Logo

Maintenance Infos

Quantification of Inter- and Intra-Tumor Heterogeneity Using Medical Imaging and Its Implication on Response to Radiotherapy in Head and Neck Cancer


Bogowicz, Marta. Quantification of Inter- and Intra-Tumor Heterogeneity Using Medical Imaging and Its Implication on Response to Radiotherapy in Head and Neck Cancer. 2017, University of Zurich, Faculty of Science.

Abstract

Cancer is a heterogeneous disease, showing intra- and inter-tumor genetic and phenotypic variability [1]. This variability translates to differential radiosensitivity and in consequence differential response to radiotherapy. Head and neck squamous cell carcinoma (HNSCC) accounts for around 5-10% of new cancer cases in developed countries [2]. It shows a heterogeneous response to radiochemotherapy with loco-regional control and 5 years overall survival ranging from below 50% to 80%. Few molecular factors were linked to outcome prognosis in HNSCC, for example human papilloma virus (HPV) infection. However, tissue-based biomarkers from tumor biopsies may not account for intra-tumor heterogeneity [3, 4]. This PhD project aims to identify new tumor phenotypes in HNSCC related to worse prognosis of treatment outcome using medical imaging techniques, which provides a 3D surrogate of tumor biology. Tumor density, metabolism and perfusion were studied in respect to different HNSCC subtypes and radiochemotherapy outcome. A quantitative and comprehensive image analysis method, radiomics, was used to link intra-tumor heterogeneity and treatment outcome. Radiomics comprises four types of descriptors: shape, intensity, texture and filter-based. It not only quantifies general properties of a tumor, for example higher metabolic activity, but also provides information about the intra-tumor heterogeneity. In the first subproject, I have analyzed tumor perfusion, metabolism and their correlation in subgroups of HNSCC based on: tumor subtype (oropharynx, hypopharynx, larynx and oral cavity), tumor stage (T1/T2 vs T3/T4) and HPV status. Computed tomography perfusion (CTP) and 18F-fludeoxyglucose positron emission tomography (18F-FDG PET), from 41 HNSCC patients were analyzed. Three perfusion parameters: blood volume (BV), blood flow (BF) and mean transit time (MTT), were computed. Difference in perfusion parameters between the gross tumor volume (GTV) and its surrounding tissue were investigated. Tumor subgroups related to worse prognosis (T3/T4 and HPV negative) showed increased BV and MTT in comparison to surrounding healthy tissue. Additionally, I have shown that the correlation of FDG uptake and perfusion is tumor subgroup dependent. I have observed positive correlation only for HPV positive (r = 0.86, p = 0.04) and oropharyngeal (r=0.63, p = 0.05) cancer. CTP consists of repeated CT scans and is thus dose intensive. I have performed a separate study using Alderson phantom to adapt our clinical CTP head and neck protocol. The endpoint was a decrease in delivered dose and maintenance of image quality. Our standard protocol on GE revolution CT is 100 kV, 80 mAs, 5 mm slice thickness and filtered back projection algorithm. I have adapted the percentage of an adaptive statistical iterative reconstruction (ASiR), slice thickness, tube current and voltage. The signal to noise ratio was measured in 7 predefined regions of interest and the effective dose was estimated using thermoluminescent dosimeters. The optimized protocol used 80 kV, a tube current adapted based on anatomy from 15 to 80 mAs, 2.5 mm slice thickness and 50% of ASiR reconstruction. The effective dose was decreased by factor of 2 whereas the image quality was maintained. In the second part of the thesis, I have investigated radiomics for its ability to predict treatment outcome and its correlation to tumor biology. An in-house radiomics software implementation was developed in Python programing language (v 2.7). Most of the radiomics studies are performed using in-house implementations or open source codes and the implemented workflows are currently not fully standardized. Therefore, I have validated my implementation against implementation from MAASTRO clinic, Maastricht, the Netherlands. I have also used both implementations to train local tumor control models based on 18F-FDG PET imaging 3 months post-radiochemotherapy (128 patients). Only 80 out of 649 radiomic features, available in both implementations and based on the same mathematical definition, were reproducible between the implementations (intraclass correlation coefficient ICC > 0.8). In the univariate Cox regression feature’s prognostic power depended strongly on used implementation. The main causes of irreproducibility were differences in contour mask creation, translation of bin size to filtered images, and type of the used transform decimated vs undecimated wavelet transform. In another radiomics robustness study I have investigated the stability of radiomic features in respect to different CTP calculation factors. Some of the CTP calculation factors are difficult to standardize (arterial input function definition and noise threshold in the calculation) and thus should be considered before linking CTP radiomics with clinical outcome. I have analyzed CTP scans in lung (n = 11) and head and neck cancer (n = 11). 255 out of 945 CTP radiomic features were stable in both tumor sites in respect to artery contouring and noise threshold. Among them, I have identified 10 groups of radiomic features, after the correction for inter-features correlations and correlation to tumor volume. These features should be further tested for their prognostic power. In the prognostic modeling, I have investigated the link between local tumor control and radiomics in HNSCC based on contrast-enhanced CT and 18F-FDG PET pre-treatment imaging. I have used two cohorts of patients: retrospective for models training (n > 90 patients) and prospective cohort from institutional phase II study with a standardized imaging protocol for models validation (n > 50 patients). I have observed that tumors more heterogeneous in CT density were at higher risk for tumor recurrence. This model had a higher prognostic power than model incorporating clinical prognostic factors (tumor stage, volume and HPV status) or combination of CT radiomics and clinical factors, concordance index (CI) in the validation cohort CIradiomics = 0.78, CIclinical = 0.73 and CIcombination = 0.76. In a follow-up study, I have investigated whether the inclusion of metabolic information can further improve radiomics-based local tumor control modeling. I have observed that round tumors (based on 18-FDG PET autosegmentation) with a focused region of high FDG uptake surrounded by a rim of low FDG uptake were linked with better prognosis. However, this model did not outperform the CT based model. In the validation cohort evaluated in this study, both models achieved CI around 0.7. Also the combination of PET and CT radiomics did not improve the predictions. Nevertheless, the PET radiomics model showed a better calibration, which may be linked to the presence of metal artifacts in CT in head and neck region. To link the abstract radiomic features with tumor biology, I have correlated CT radiomics with HPV status. I have observed that tumors more homogenous in CT density tend to be HPV positive. Although, this signature (set of radiomic features) has a similar interpretation to local tumor control signature, it comprised different features and the signatures were not correlated with each other. For example local tumor control CT radiomics model was also prognostic in HPV negative subgroup of patients. In summary, I have shown that biological information can be recovered even from simple morphological imaging (CT). Additionally, I have identified imaging signatures, based on differences in perfusion between tumor and its surrounding as well as CT and PET radiomics, which were linked with worse outcome prognosis. These signatures need to be further validated in an external cohort of patients and treatment intensification options for worse prognosis groups have to be defined.

Abstract

Cancer is a heterogeneous disease, showing intra- and inter-tumor genetic and phenotypic variability [1]. This variability translates to differential radiosensitivity and in consequence differential response to radiotherapy. Head and neck squamous cell carcinoma (HNSCC) accounts for around 5-10% of new cancer cases in developed countries [2]. It shows a heterogeneous response to radiochemotherapy with loco-regional control and 5 years overall survival ranging from below 50% to 80%. Few molecular factors were linked to outcome prognosis in HNSCC, for example human papilloma virus (HPV) infection. However, tissue-based biomarkers from tumor biopsies may not account for intra-tumor heterogeneity [3, 4]. This PhD project aims to identify new tumor phenotypes in HNSCC related to worse prognosis of treatment outcome using medical imaging techniques, which provides a 3D surrogate of tumor biology. Tumor density, metabolism and perfusion were studied in respect to different HNSCC subtypes and radiochemotherapy outcome. A quantitative and comprehensive image analysis method, radiomics, was used to link intra-tumor heterogeneity and treatment outcome. Radiomics comprises four types of descriptors: shape, intensity, texture and filter-based. It not only quantifies general properties of a tumor, for example higher metabolic activity, but also provides information about the intra-tumor heterogeneity. In the first subproject, I have analyzed tumor perfusion, metabolism and their correlation in subgroups of HNSCC based on: tumor subtype (oropharynx, hypopharynx, larynx and oral cavity), tumor stage (T1/T2 vs T3/T4) and HPV status. Computed tomography perfusion (CTP) and 18F-fludeoxyglucose positron emission tomography (18F-FDG PET), from 41 HNSCC patients were analyzed. Three perfusion parameters: blood volume (BV), blood flow (BF) and mean transit time (MTT), were computed. Difference in perfusion parameters between the gross tumor volume (GTV) and its surrounding tissue were investigated. Tumor subgroups related to worse prognosis (T3/T4 and HPV negative) showed increased BV and MTT in comparison to surrounding healthy tissue. Additionally, I have shown that the correlation of FDG uptake and perfusion is tumor subgroup dependent. I have observed positive correlation only for HPV positive (r = 0.86, p = 0.04) and oropharyngeal (r=0.63, p = 0.05) cancer. CTP consists of repeated CT scans and is thus dose intensive. I have performed a separate study using Alderson phantom to adapt our clinical CTP head and neck protocol. The endpoint was a decrease in delivered dose and maintenance of image quality. Our standard protocol on GE revolution CT is 100 kV, 80 mAs, 5 mm slice thickness and filtered back projection algorithm. I have adapted the percentage of an adaptive statistical iterative reconstruction (ASiR), slice thickness, tube current and voltage. The signal to noise ratio was measured in 7 predefined regions of interest and the effective dose was estimated using thermoluminescent dosimeters. The optimized protocol used 80 kV, a tube current adapted based on anatomy from 15 to 80 mAs, 2.5 mm slice thickness and 50% of ASiR reconstruction. The effective dose was decreased by factor of 2 whereas the image quality was maintained. In the second part of the thesis, I have investigated radiomics for its ability to predict treatment outcome and its correlation to tumor biology. An in-house radiomics software implementation was developed in Python programing language (v 2.7). Most of the radiomics studies are performed using in-house implementations or open source codes and the implemented workflows are currently not fully standardized. Therefore, I have validated my implementation against implementation from MAASTRO clinic, Maastricht, the Netherlands. I have also used both implementations to train local tumor control models based on 18F-FDG PET imaging 3 months post-radiochemotherapy (128 patients). Only 80 out of 649 radiomic features, available in both implementations and based on the same mathematical definition, were reproducible between the implementations (intraclass correlation coefficient ICC > 0.8). In the univariate Cox regression feature’s prognostic power depended strongly on used implementation. The main causes of irreproducibility were differences in contour mask creation, translation of bin size to filtered images, and type of the used transform decimated vs undecimated wavelet transform. In another radiomics robustness study I have investigated the stability of radiomic features in respect to different CTP calculation factors. Some of the CTP calculation factors are difficult to standardize (arterial input function definition and noise threshold in the calculation) and thus should be considered before linking CTP radiomics with clinical outcome. I have analyzed CTP scans in lung (n = 11) and head and neck cancer (n = 11). 255 out of 945 CTP radiomic features were stable in both tumor sites in respect to artery contouring and noise threshold. Among them, I have identified 10 groups of radiomic features, after the correction for inter-features correlations and correlation to tumor volume. These features should be further tested for their prognostic power. In the prognostic modeling, I have investigated the link between local tumor control and radiomics in HNSCC based on contrast-enhanced CT and 18F-FDG PET pre-treatment imaging. I have used two cohorts of patients: retrospective for models training (n > 90 patients) and prospective cohort from institutional phase II study with a standardized imaging protocol for models validation (n > 50 patients). I have observed that tumors more heterogeneous in CT density were at higher risk for tumor recurrence. This model had a higher prognostic power than model incorporating clinical prognostic factors (tumor stage, volume and HPV status) or combination of CT radiomics and clinical factors, concordance index (CI) in the validation cohort CIradiomics = 0.78, CIclinical = 0.73 and CIcombination = 0.76. In a follow-up study, I have investigated whether the inclusion of metabolic information can further improve radiomics-based local tumor control modeling. I have observed that round tumors (based on 18-FDG PET autosegmentation) with a focused region of high FDG uptake surrounded by a rim of low FDG uptake were linked with better prognosis. However, this model did not outperform the CT based model. In the validation cohort evaluated in this study, both models achieved CI around 0.7. Also the combination of PET and CT radiomics did not improve the predictions. Nevertheless, the PET radiomics model showed a better calibration, which may be linked to the presence of metal artifacts in CT in head and neck region. To link the abstract radiomic features with tumor biology, I have correlated CT radiomics with HPV status. I have observed that tumors more homogenous in CT density tend to be HPV positive. Although, this signature (set of radiomic features) has a similar interpretation to local tumor control signature, it comprised different features and the signatures were not correlated with each other. For example local tumor control CT radiomics model was also prognostic in HPV negative subgroup of patients. In summary, I have shown that biological information can be recovered even from simple morphological imaging (CT). Additionally, I have identified imaging signatures, based on differences in perfusion between tumor and its surrounding as well as CT and PET radiomics, which were linked with worse outcome prognosis. These signatures need to be further validated in an external cohort of patients and treatment intensification options for worse prognosis groups have to be defined.

Statistics

Downloads

34 downloads since deposited on 14 Feb 2019
26 downloads since 12 months
Detailed statistics

Additional indexing

Item Type:Dissertation (monographical)
Referees:Guckenberger Matthias, Riesterer Oliver, Schneider Uwe, Tanadini-Lang Stephanie, Unkelbach Jan
Communities & Collections:04 Faculty of Medicine > University Hospital Zurich > Clinic for Radiation Oncology
UZH Dissertations
Dewey Decimal Classification:610 Medicine & health
Language:English
Place of Publication:Zürich
Date:2017
Deposited On:14 Feb 2019 09:02
Last Modified:15 Apr 2021 15:05
OA Status:Green

Download

Green Open Access

Download PDF  'Quantification of Inter- and Intra-Tumor Heterogeneity Using Medical Imaging and Its Implication on Response to Radiotherapy in Head and Neck Cancer'.
Preview
Content: Published Version
Filetype: PDF
Size: 8MB