Next Article in Journal
Exploratory Analysis of Lenvatinib Therapy in Patients with Unresectable Hepatocellular Carcinoma Who Have Failed Prior PD−1/PD-L1 Checkpoint Blockade
Next Article in Special Issue
Complex Mandibular Reconstruction for Head and Neck Squamous Cell Carcinoma—The Ongoing Challenge in Reconstruction and Rehabilitation
Previous Article in Journal
Rare Germline Genetic Variants and the Risks of Epithelial Ovarian Cancer
Previous Article in Special Issue
Cell-Free HPV-DNA as a Biomarker for Oropharyngeal Squamous Cell Carcinoma—A Step Towards Personalized Medicine?
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Comprehensive Analysis of Tumour Sub-Volumes for Radiomic Risk Modelling in Locally Advanced HNSCC

by Stefan Leger 1,2,3,*, Alex Zwanenburg 1,2,3, Karoline Leger 1,2,3,4, Fabian Lohaus 1,2,3,4, Annett Linge 1,2,3,4, Andreas Schreiber 5, Goda Kalinauskaite 6,7, Inge Tinhofer 6,7, Nika Guberina 8,9, Maja Guberina 8,9, Panagiotis Balermpas 10,11, Jens von der Grün 10,11, Ute Ganswindt 12,13,14,15, Claus Belka 12,13,14, Jan C. Peeken 12,16,17, Stephanie E. Combs 12,16,17, Simon Boeke 18,19, Daniel Zips 18,19, Christian Richter 1,2,4,20, Mechthild Krause 1,2,3,4,20, Michael Baumann 1,2,3,4,20,21, Esther G.C. Troost 1,2,3,4,20,† and Steffen Löck 1,2,3,†add Show full author list remove Hide full author list
1
OncoRay—National Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universität Dresden, Helmholtz-Zentrum Dresden—Rossendorf, 01307 Dresden, Germany
2
German Cancer Research Center (DKFZ), Heidelberg and German Cancer Consortium (DKTK) Partner Site, 01307 Dresden, Germany
3
National Center for Tumor Diseases (NCT), Partner Site Dresden of the German Cancer Research Center (DKFZ), Faculty of Medicine and University Hospital Carl Gustav Carus and Technische Universität Dresden, 01307 Dresden, Germany
4
Department of Radiotherapy and Radiation Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universität Dresden, 01307 Dresden, Germany
5
Department of Radiotherapy, Hospital Dresden-Friedrichstadt, 01067 Dresden, Germany
6
German Cancer Research Center (DKFZ), Heidelberg and German Cancer Consortium (DKTK) Partner Site, 10117 Berlin, Germany
7
Department of Radiooncology and Radiotherapy, Charité University Hospital, 10117 Berlin, Germany
8
German Cancer Research Center (DKFZ), Heidelberg and German Cancer Consortium (DKTK) Partner Site, 45147 Essen, Germany
9
Department of Radiotherapy, University Hospital Essen, Medical Faculty, University of Duisburg-Essen, 45147 Essen, Germany
10
German Cancer Research Center (DKFZ), Heidelberg and German Cancer Consortium (DKTK) Partner Site, 60596 Frankfurt, Germany
11
Department of Radiotherapy and Oncology, Goethe-University Frankfurt, 60596 Frankfurt, Germany
12
German Cancer Research Center (DKFZ), Heidelberg and German Cancer Consortium (DKTK) Partner Site, 81377 Munich, Germany
13
Department of Radiation Oncology, Ludwig-Maximilians-Universität, 81377 Munich, Germany
14
Clinical Cooperation Group, Personalized Radiotherapy in Head and Neck Cancer, Helmholtz Zentrum, 81377 Munich, Germany
15
Department of Radiation Oncology, Medical University of Innsbruck, Anichstraße 35, A-6020 Innsbruck, Austria
16
Department of Radiation Oncology, Technische Universität München, 81675 Munich, Germany
17
Institute of Radiation Medicine (IRM), Helmholtz Zentrum München, 85764 Neuherberg, Germany
18
German Cancer Research Center (DKFZ), Heidelberg and German Cancer Consortium (DKTK) Partner Site, 72076 Tübingen, Germany
19
Department of Radiation Oncology, Faculty of Medicine and University Hospital Tübingen, Eberhard Karls Universität Tübingen, 72076 Tübingen, Germany
20
Institute of Radiooncology—OncoRay, Helmholtz-Zentrum Dresden—Rossendorf, 01328 Dresden, Germany
21
German Cancer Research Center (DKFZ), 69120 Heidelberg, Germany
*
Author to whom correspondence should be addressed.
These authors share senior authorship.
Submission received: 28 August 2020 / Revised: 7 October 2020 / Accepted: 13 October 2020 / Published: 19 October 2020
(This article belongs to the Special Issue Advances in Head and Neck Squamous Cell Carcinoma (HNSCC))

Abstract

:

Simple Summary

Radiomic risk models are usually based on imaging features, which are extracted from the entire gross tumour volume (GTV entire ). This approach does not explicitly consider the complex biological structure of the tumours. Therefore, in this retrospective study, we investigated the prognostic value of radiomic analyses based on different tumour sub-volumes using computed tomography imaging of patients with locally advanced head and neck squamous cell carcinoma who were treated with primary radio-chemotherapy. The GTV entire was cropped by different margins to define the rim and corresponding core sub-volumes of the tumour. Furthermore, the best performing tumour rim sub-volume was extended into surrounding tissue with different margins. As a result, the models based on the 5 mm tumour rim and on the 3 mm extended rim sub-volume showed an improved performance compared to models based on the corresponding tumour core. This indicates that the consideration of tumour sub-volumes may help to improve radiomic risk models.

Abstract

Imaging features for radiomic analyses are commonly calculated from the entire gross tumour volume (GTV entire ). However, tumours are biologically complex and the consideration of different tumour regions in radiomic models may lead to an improved outcome prediction. Therefore, we investigated the prognostic value of radiomic analyses based on different tumour sub-volumes using computed tomography imaging of patients with locally advanced head and neck squamous cell carcinoma. The GTV entire was cropped by different margins to define the rim and the corresponding core sub-volumes of the tumour. Subsequently, the best performing tumour rim sub-volume was extended into surrounding tissue with different margins. Radiomic risk models were developed and validated using a retrospective cohort consisting of 291 patients in one of the six Partner Sites of the German Cancer Consortium Radiation Oncology Group treated between 2005 and 2013. The validation concordance index (C-index) averaged over all applied learning algorithms and feature selection methods using the GTV entire achieved a moderate prognostic performance for loco-regional tumour control (C-index: 0.61 ± 0.04 (mean ± std)). The models based on the 5 mm tumour rim and on the 3 mm extended rim sub-volume showed higher median performances (C-index: 0.65 ± 0.02 and 0.64 ± 0.05, respectively), while models based on the corresponding tumour core volumes performed less (C-index: 0.59 ± 0.01). The difference in C-index between the 5 mm tumour rim and the corresponding core volume showed a statistical trend (p = 0.10). After additional prospective validation, the consideration of tumour sub-volumes may be a promising way to improve prognostic radiomic risk models.

1. Introduction

The individualisation of radiation oncology is a major objective in modern cancer therapy [1]. Radiomics aims to characterise the tumour phenotype using advanced image features to predict patient-specific outcome. Commonly, radiomic features are computed and extracted using the entire gross tumour volume (GTV entire ) [2,3,4,5]. Such an approach assumes that the individual tumour is either homogeneous or heterogeneous, but uniformly distributed over the entire tumour volume. However, tumours are biologically complex and exhibit substantial spatial variation, e.g., in gene expression and in microscopic structure [6]. Such spatial variation may be caused by, e.g., hypoxia and necrosis which may appear in the tumour core and high cell proliferation and infiltrating tumour cell growth, which may occur at the tumour periphery [7]. Some of these regional tumour variations are apparent in imaging data, e.g., necrosis or tumour vascularisation detected by magnetic resonance imaging (MRI) or tumour hypoxia measured by 18 F-fluoromisonidazole positron emission tomography (FMISO-PET) [8,9,10,11,12]. Furthermore, different regions within an individual tumour may differ in radio-sensitivity, which may depend on the distribution of cancer stem cells and localised genetic or molecular alterations [13,14]. In the case of head and neck squamous cell carcinoma (HNSCC), several studies have shown that the tumour micro-environment plays a major role for cancer development and progression [15]. Alsahafi et al. [16] showed that the poor response to therapy and the aggressive nature of HNSCC are not only caused by the complex alterations in intracellular signalling pathways, but are also influenced by the behaviour of the extracellular micro-environment. As consequences, such spatial variations may affect the performance of image-based risk models.
Thus far, only few studies have investigated and analysed specific tumour sub-volumes for radiomic risk modelling. Recently, Algohary et al. [17] showed that the combination of peri-tumoural and intra-tumoural radiomic features derived from prostate bi-parametric MR images leads to an improved risk assessment of prostate cancer patients. Furthermore, Grove et al. [18] showed that the expressions of 2-dimensional radiomic features computed on the rim of the tumour differed from those calculated on the tumour core. The ratio of tumour rim and core features led to an improved prediction of overall survival (OS) in non-small cell lung cancer patients. Wu et al. [6] identified clinically relevant tumour sub-volumes to characterise the regional heterogeneity of tumours in breast cancer patients based on dynamic contrast enhanced magnetic resonance imaging. The resulting risk models based on the identified sub-volumes also showed an improved outcome prediction compared to models based on the GTV entire . In a further study, Wu et al. [19] identified different tumour sub-volumes using computed tomography (CT) and 18 F-fluorodeoxyglucose PET (FDG-PET) imaging of lung cancer patients. It was shown that spatially distinct sub-volumes are linked to higher risk of recurrence compared to the GTV entire , resulting in an improved model prediction of OS.
Aside from these initial findings, in most of the previously described studies, only individual clinical parameters or radiomic features (e.g., tumour volume) were investigated. Therefore, systematic investigations of the potential of radiomic risk models based on different tumour sub-volumes are still sparse.
In the present study, we systematically compared radiomic models based on two different sub-volumes of the GTV entire , the outer tumour rim and the complementary tumour core, using pre-treatment CT imaging [20,21]. A multi-centre cohort of 291 patients with locally advanced HNSCC treated by primary radio-chemotherapy was considered. For the prediction of loco-regional tumour control (LRC), risk models were developed and independently validated. Patients were stratified into groups at low and high risk of loco-regional recurrence. Furthermore, we investigated the prognostic performance of the developed models for sub-groups of small and large tumours and extended the tumour rim beyond the GTV entire to account for potential sub-microscopic spread, leading to the clinical target volume [22].

2. Materials and Methods

2.1. Characteristics of Patient Cohorts

A retrospective multi-centre cohort consisting of 291 patients with histologically confirmed loco-regionally advanced HNSCC was used. All patients received primary radio-chemotherapy (RCT) and underwent a CT scan with or without contrast-enhancement for treatment-planning purpose. The multi-centre cohort was divided into an exploratory and a validation cohort by an approximate ratio of 2:1. In the exploratory cohort, 149 of the 206 patients were treated in one of the six Partner Sites of the German Cancer Consortium Radiation Oncology Group (DKTK-ROG) between 2005 and 2011 [23]. The remaining 57 patients were treated at the University Hospital Dresden (UKD) between 1999 and 2006. The validation cohort consisted of 85 patients from which 51 patients received their treatment within a prospective clinical trial (ClinicalTrials.gov Identifier: NCT00180180) at the UKD between 2006 and 2012 [9,12]. The remaining 34 patients were treated at the UKD or the Radiotherapy Centre Dresden-Friedrichstadt between 2005 and 2009 as well as at the University Hospital Tübingen between 2008 and 2013. Patient characteristics for the exploratory and the independent validation cohort are summarised in Table 1.
Radiomic risk models were developed to predict the primary clinical endpoint LRC, which was defined as the time from the first day of RCT to the date of loco-regional recurrence (event) or to the end of follow-up (censoring). Ethical approval for the multi-centre retrospective analyses of clinical and imaging data was obtained from the Ethics Committee at the Technische Universität Dresden (EK177042017, May 2017). All analyses were carried out in accordance with the relevant guidelines and regulations. Informed consent was obtained from all patients.

2.2. Tumour Sub-Volume Definition and Feature Computation

The analysis was divided into two subsequent steps, which are shown in Figure 1. The GTV entire , i.e., the primary gross tumour volume, was manually delineated on each planning CT scan by a radiation oncologist using the CT image information only. Subsequently, the image voxel spacing was resampled using cubic spline image interpolation to an isotropic voxel size of 1.0 × 1.0 × 1.0 mm 3 to correct for differences in voxel spacing and slice thickness between the cohorts [2,24].
Based on the delineated GTV entire , two distinct sub-volumes were generated. The outer contour of the GTV entire was cropped by different margins (3 and 5 mm) to define the rim of the tumour (GTV 3mm-rim and GTV 5mm-rim , respectively). The corresponding remaining sub-volumes were defined as tumour core (GTV 3mm-core and GTV 5mm-core , respectively). The minimum core volume was restricted to 40% of the entire tumour volume to avoid disappearance of the core sub-volume in small tumours. Furthermore, the best performing tumour rim sub-volume was extended (GTV rim+ext ) into surrounding tissue with different distances (1, 2, 3 and 5 mm) to assess the prognostic performance of the microscopic tumour extension.
Nine additional images were created by applying spatial filtering to the base image to emphasise image characteristics, such as edges and blobs. Eight additional images were created by applying a stationary coiflet-1 wavelet high-/low-pass filter along each of the three spatial dimensions. One further image was created by applying a Laplacian of Gaussian (LoG) filter consisting of five different filter kernel widths (1.0, 2.0, 3.0, 5.0 and 6.0 mm). Subsequently, the tumour mask was re-segmented to include only soft tissue voxels between −150 and 180 Hounsfield units, thereby removing voxels containing air or bone, which may affect feature expression. Features were implemented in compliance with the Image Biomarker Standardisation Initiative [25]. A total of 1538 features were computed and extracted from each sub-volume. A total of 18 statistical, 38 histogram-based and 95 texture features were calculated on the base image and on the nine transformed images. Moreover, 28 morphological features were determined on the base image only. The configuration settings for the image feature computation and extraction is summarised in Table A2.

2.3. Radiomic Risk Modelling

Radiomic risk models were developed using an end-to-end modelling framework, which consists of five steps: (I) feature pre-processing; (II) feature selection; (III) hyper-parameter optimisation; (IV) model development; and (V) model validation. The risk models were generated as previously described [4]. Briefly, after feature normalisation and clustering, feature selection was performed multiple times using 1000 bootstrap samples of the exploratory cohort. Subsequently, model training was conducted on 1000 bootstrap samples of the exploratory cohort, using the highest ranked features as well as the optimised hyper-parameter set. Finally, an ensemble prediction was made by averaging the predicted risk scores of each model for both the exploratory and the independent validation cohort separately.
Combinations of five feature selection methods and six learning algorithms were used for model development to reduce the risk of incidental findings, based on the recommendation in Leger et al. [4]. The following feature selection methods were used: Spearman correlation, mutual information maximisation (MIM), mutual information feature selection (MIFS), minimum redundancy maximum relevance (MRMR) and random forest variable importance (RFVI). The six learning algorithms comprised: Cox proportional hazard model (Cox), boosted tree-Cox (BT-Cox), boosted gradient linear model-Cox (BGLM-Cox), random survival forest (RSF) and maximally selected rank statistics random forest (MSR-RF) as well as the full-parametric BT-Weibull model. Table A3 summarises the definition of the hyper-parameters of the feature selection methods and of the machine learning algorithms, which were used during the hyper-parameter optimisation.

2.4. Performance Assessments

(I) The prognostic performance of the radiomic models was assessed on the exploratory and on the independent validation cohort using the concordance index (C-index) [26,27]. The C-index is a generalisation of the area under the curve for continuous time-to-event survival data and a C-index of 0.5 describes a random prediction, whereas a perfectly predicting model has C-index of 1.0. Risk models were developed based on GTV entire , GTV 3mm-rim and GTV 5mm-rim , as well as the corresponding core volumes GTV 3mm-core and GTV 5mm-core .
The median C-indices over all combination of feature selection methods and machine learning algorithms were determined based on the exploratory and the validation cohort for each tumour sub-volume. For the considered feature selection methods and machine learning algorithms, the model performance was statistically compared between the GTV 3mm-rim and GTV 5mm-rim and their corresponding core volumes using a multi-level model approach (MLA), which is described in Appendix A.1 [5]. Subsequently, representative model combinations for each sub-volume were selected, consisting of one feature section method and one learning algorithm. To choose this representative model, the median performances of every feature selection method over all learning algorithms and vice versa were determined. The model generated by the feature selection method with a C-index closest to the median feature selection performances and the learning algorithm with a C-index closest to the median learning algorithm performances was then selected. For the further analyses, the representative models based on the sub-volume of (a) the GTV entire ; (b) the tumour rim; (c) the corresponding core and (d) the extended rim were investigated in more detail. In addition, we assigned the patients of the validation cohort into two sub-groups according their initial tumour volume using 20 cm 3 as a threshold value, which corresponds to a tumour radius of approximately 1.5 cm in the case of a spherical tumour. Subsequently, we investigated the prognostic performance of the developed models using the resulting sub-groups individually.
(II) Risk-based patient stratification into groups at low and high risk of loco-regional recurrence was performed for each tumour sub-volume and for all model combinations. The results for the selected models (a)–(d) were analysed in more detail. Patients were stratified into a low and high risk group based on the predicted risk of the radiomics models. The cut-off value used for stratification was based on the median predicted risk (median risk ) determined on the exploratory cohort. This cut-off value was directly applied to the validation cohort. Survival curves were estimated using the Kaplan–Meier method and the stratification was compared using log-rank tests. Log-rank test p-values < 0.05 were considered to be statistically significant.
(III) Radiomic signatures were analysed in detail for the models trained on the different selected tumour sub-volumes (a)–(d). Features included in the signatures and their expression values were depicted as heatmaps for the exploratory and the validation cohort. For this purpose, all patients were sorted according to their predicted risk and to their risk group stratification. To quantify the overall importance of the identified features, univariate significance of the individual radiomic features included in the signatures were tested by the Cox model on the entire patient cohort.

3. Results

The number of loco-regional recurrences was 84 for the exploratory and 28 for the independent validation cohort, respectively. The primary endpoint LRC showed no significant difference between both cohorts (p = 0.26). The median follow-up time was 21.2 months (range: 1.2–131.9 months) for the exploratory and 24.3 months (range: 1.3–107.2 months) for the validation cohort (p = 0.64). The median GTV entire was 29.2 cm 3 (range: 4.4–322.2 cm 3 ) in the exploratory cohort and 40.1 cm 3 (range: 2.7–239.0 cm 3 ) in the validation cohort (p = 0.067, Table 1). For further analyses, 3 mm and 5 mm margins were subtracted from the GTV entire , respectively. The median volume fractions of the resulting tumour rim sub-volumes were 47.7% (range: 26.1–59.5%) for the GTV 3mm-rim and 52.8% (range: 33.1–59.9%) for the GTV 5mm-rim sub-volumes in the exploratory cohort and 46.4% (range: 26.1–59.5%) and 52.1% (range: 33.1–59.9%) in the validation cohort (p = 0.066 and p = 0.081, respectively).

3.1. Prognostic Performance

Radiomic models were developed and validated based on the different (sub-)volumes of the tumour. Their performance for the prognosis of LRC is summarised in Table 2 for both cohorts. The median C-index on the exploratory cohort was between 0.72 and 0.76 for the considered sub-volumes. For the validation cohort, models based on the GTV entire achieved a median prognostic performance of 0.61 ± 0.04 (median ± standard deviation (SD)), while the models based on the tumour rim sub-volumes showed a slightly better median performance on the validation cohort (C-index: GTV 3mm-rim : 0.63 ± 0.03 and GTV 5mm-rim : 0.65 ± 0.02, respectively). The core-based risk models revealed the lowest prognostic performance on the validation cohort (C-index: GTV 3mm-core : 0.60 ± 0.02 and GTV 5mm-core : 0.59 ± 0.01, respectively). The difference in C-index between GTV 5mm-rim and GTV 5mm-core showed a statistical trend (MLA: p = 0.10), while the difference between GTV 3mm-rim and GTV 3mm-core was not statistically significant (MLA: p = 0.50). The median performances of the sub-group analyses showed similar results for small GTV entire (≤20 cm 3 ) between the tumour rim- and core-based models (3 mm: 0.62 vs. 0.63 and 5 mm: 0.67 vs. 0.69, respectively), whereas, the differences between rim and core were larger for larger GTV entire (3 mm: 0.61 vs. 0.57 and 5 mm: 0.61 vs. 0.57) in the validation cohort. Furthermore, overall performance was higher for the sub-group of smaller tumours.
The C-indices of the representative models for each tumour sub-volume are shown in Table 3. Among all GTV entire -based risk models, the RSF algorithm in combination with the RFVI feature selection method was selected as representative model for further analysis (C-index: 0.75, 95% confidence interval [0.71–0.81]). On the validation cohort, this model achieved a C-index of 0.63 ([0.49–0.67]). The RSF–MIM model trained on the GTV 5mm-rim was selected as representative model compared to all other rim-based models on the exploratory cohort (C-index: 0.77, [0.72–0.82]). This model attained an improved performance on the validation cohort (C-index: 0.66, [0.52–0.69]), which was slightly higher compared to the GTV entire -based model. The corresponding GTV 5mm-core -based model (RSF–MIM) showed a lower prognostic performance on the validation cohort (C-index: 0.61, [0.49–0.69]) compared to the GTV 5mm-rim model. Figure A1 and Figure A2 show the prognostic performance for the considered feature selection methods and learning algorithms based on the GTV entire , GTV 3mm-rim and GTV 5mm-rim , as well as the corresponding core sub-volumes on the exploratory and the validation cohorts.
The GTV 5mm-rim sub-volume, which achieved the highest prognostic performance among all rim-based models was subsequently extended by different margins beyond the originally delineated tumour into surrounding tissue. The tumour extensions GTV 5mm-rim+2mm-ext and GTV 5mm-rim+3mm-ext showed the highest median performances on the validation cohort (C-indices: 0.63 ± 0.03 and 0.64 ± 0.05, respectively). The C-index of the representative models trained on the different tumour extensions are shown in Table 2. The representative model trained on the GTV 5mm-rim+3mm-ext (RSF–RFVI) achieved a slightly better performance in validation compared to the model based on the GTV entire and on the GTV 5mm-rim (C-index: 0.67, [0.60–0.77]). The resulting C-indices for all extended sub-volumes and developed radiomic risk models on the exploratory and validation cohort are summarised in Figure A3 and Figure A4, respectively. The hyper-parameters including the optimised values for the representative models are given in Table A4.

3.2. Risk-Based Patient Stratification

Patients were stratified into groups at low and high risk for loco-regional recurrence based on the model prediction of the exploratory cohort. Table 3 shows the p-values of the log-rank test for LRC for all representative models on the validation cohort based on the median risk cut-off. The RSF–RFVI model trained on GTV entire was able to stratify patients into low and high risk groups with a significant difference in LRC (p = 0.012). A slightly improved stratification could be achieved by the GTV 3mm-rim - and GTV 5mm-rim -based models (p = 0.005 and p = 0.006, respectively) as well as by the extended volume GTV 5mm-rim+3mm-ext (p < 0.001). Stratification based on the predicted risk of the corresponding GTV 5mm-core model (RSF–MIM) did not lead to significant differences in LRC between both groups (p = 0.11).
Figure 2 shows the Kaplan–Meier curves using the median risk cut-off for the representative models based on (a) GTV entire , (b) GTV 5mm-rim , (c) GTV 5mm-core and (d) GTV 5mm-rim+3mm-ext , respectively, for the validation cohort. The resulting p-values for all considered sub-volumes and developed radiomic risk models on the validation cohort are summarised in Figure A5 and Figure A6.

3.3. Radiomic Signature Analysis

Radiomics signatures were investigated for the representative models based on GTV entire , GTV 5mm-rim , GTV 5mm-core and GTV 5mm-rim+3mm-ext . Figure 3 shows the feature expressions of the developed signatures for each patient in a heatmap. Image features within the signatures are listed in Table A5.
The developed signatures for the different models (a)–(d) consist of two to ten imaging features extracted from the original and wavelet transformed images. The selected features typically comprise first-order statistical or texture-based features. For instance, the ’statistics energy’ feature, which describes the overall density of the tumour volume, appears in all four signatures as a single or as a feature within a cluster [2]. Furthermore, the signatures of the trained models (a) and (b) consist of the same intensity-volume histogram feature (i.e., ‘ivh_diff_v10_v90’) computed and extracted from the wavelet transformed images. This feature describes the difference between the largest volume fractions at two different intensity values of at least 10% and 90% [25,28]. The selected features for (a) and (b) were mostly based on the low-pass wavelet transformed images, which may contain reduced noise. Features within the signatures (c) and (d) were mostly computed on the high-pass wavelet transformed images, which may characterise edges and blobs within the considered regions. For all developed signatures (a)–(d), almost all features were significantly associated with LRC based on univariate Cox analyses using the entire patient cohort.

4. Discussion

Tumours may contain biologically complex structures and exhibit substantial spatial variation. Thus, the main objective of this study was to compare radiomic models based on different sub-volumes of the tumour, i.e., on the tumour rim, the tumour core and the macroscopic tumour extensions, in order to identify potential regions containing the most relevant prognostic information for LRC. Using CT imaging of patients with locally advanced HNSCC revealed that radiomic risk models based on tumour rim sub-volumes achieved a slightly improved prognostic performance and better patient stratification compared to models based on the corresponding core regions. Furthermore, sub-group analyses showed that the differences in prognostic performance between rim and core regions were larger for large tumours compared to small tumours. In general, our analysis showed a good median performance and a better patient stratification for the models based on the GTV 5mm-rim , while the corresponding core-based models performed slightly less. This may indicate that the tumour rim contains more prognostic information. The statistical comparison between both sub volumes led to a borderline statistical trend (MLA: p = 0.10), i.e., the presented findings require additional validation in the future.
These results are in-line with previously published data of other tumour entities [10,11,29]. For example, Dou et al. [30] showed that models based on CT-imaging features of the 3 mm rim of the GTV lead to an improved prediction of distant metastasis compared to the model based on the entire GTV for patients with locally advanced non-small cell lung cancer (NSCLC). Furthermore, Hosney et al. [31] developed a deep learning-based prediction model using a 3D convolutional neuronal network for the prediction of OS for NSCLC patients and observed that the network tended to focus on the interface between the tumour and stroma (parenchyma or pleura) regions in the CT images. In contrast to that, Keek et al. [32] concluded that the consideration of the tumour rim did not lead to an improved prediction of overall survival, loco-regional recurrence and distant metastases in stage III and IV HNSCC patients. However, for the prediction of loco-regional recurrence, a better prediction could be observed for the 5 mm rim-based model compared to the model using the GTV entire in the exploratory and validation cohort (C-index: 0.86/0.59 and 0.81/0.52, respectively). In addition, Grove et al. [18] showed that tumour-rim-based radiomic features (i.e., entropy) were higher expressed compared to features extracted from corresponding tumour-core sub-volumes in NSCLC patients. While, the entropy feature and their ratios of core and rim regions were associated with overall survival in the exploratory cohort, but not in the independent validation cohort.
The biological characteristics of the tumour rim were already discussed by published data from the Danish Head and Neck Cancer (DAHANCA) group [33]. Based on experience from pathological examination of surgical resections, the DAHANCA group concluded that for primary tumours, the risk of sub-clinical microscopic spread was around 50% of which more than 99% was within 5 mm and 95% within 4 mm of the rim of the primary tumour. Furthermore, Apolle et al. [22] showed that most solid tumours exhibit microscopic tumour extension in particular for head and neck cancer. Our findings suggest that biological processes such as microscopic spread capacity are associated with macroscopic CT imaging.
Defining the precise extent of the macroscopic tumour prior to and during RCT is difficult, especially using CT imaging without contrast enhancement [34]. Slight extensions of the delineated tumour volume into normal tissue did not reduce the performance of the radiomic risk models, which indicates that these regions may also contain prognostic information. In addition, slight extensions of the tumour may be useful for assessing feature stability, simulating different tumour delineations of different observers [35]. Furthermore, uncertainties in the delineation of the GTV entire may affect radiomic features and in turn the results of radiomic analyses. In the current study, the GTV entire was manually delineated by one expert radiation oncologist. The consideration of multiple tumour delineations of different experts or the usage of semi-automatic segmentation algorithms as well as contour randomisation techniques may help to increase the robustness of the radiomics features and improve the corresponding risk model performance, which should be investigated in the future [35,36,37].
The presented study is motivated form the assumption that hypoxic or necrotic regions preferantly appear in the tumour core due to inadequate vascular supply, and that proliferating cancer cells mainly occur in the tumour periphery [38]. Our retrospective patient cohort contains tumours with a wide range of different volumes. While necrotic and hypoxic regions will be minimal in small tumours they may be substantial in larger tumours, i.e., the prognostic value of tumour core and rim may change depending on the tumour volume. Therefore, we performed a subgroup analysis considering patients with small and large tumours separately. We found larger differences in prognostic performance between rim and core regions for larger tumours in validation, supporting this hypothesis. Still, the inclusion of patients with small and large tumours in our main analyses may affect the difference in performance between the rim- and core-based risk models. Moreover, necrotic/hypoxic regions may be heterogeneously distributed in the tumour and not be sufficiently captured by our simple approach of defining the tumour core and rim, which in addition does not consider other complex spatial and temporal variations in the tumour micro-environment. This may lead to smaller observable differences in the performance between the rim- and core-based models [14,39]. The identification and incorporation of tumour specific regional variations by more sophisticated image analysis techniques may help to overcome this gap. For instance, differential information from multi-modal imaging data, such as PET-CT or functional MRI may be used. Moreover, super-voxel algorithms can be applied to group voxels into super-voxel segments based on their grey value, e.g., using the FDG uptake value [40]. Subsequently, the resulting super-voxel segments can be further merged to generate tumour sub-volumes, e.g., by hierarchical or fuzzy c-means clustering algorithms across the entire patient cohort. Wu et al. [19] proposed such a two-stage clustering process, for the identification and determination of sub-volumes based on CT imaging combined with FDG-PET scans in lung cancer patients. Furthermore, the consideration of regions with temporal changes, e.g., due to RCT-induced tumour shrinkage or re-oxygenation using in-treatment CT images in combination with functional imaging may offer the potential to enhance radiomic risk models in future [5,39]. However, due to missing functional imaging, it was not possible to use such imaging data in this study.
In addition to radiomic features, clinical parameters may be relevant for the prediction of treatment outcome. On our cohort, from the parameters shown in Table 1, only the primary tumour volume and the derived tumour sub-volumes were significantly related to LRC (p < 0.01). These parameters revealed C-indices between 0.62 and 0.63 in the validation cohort using univariable Cox regression model. This was slightly lower than observed for the presented radiomic models based on the GTV 5mm-rim and the GTV 5mm-rim+3mm-ext (C-index: 0.65 and 0.67, respectively). While the radiomic signature based on the GTV 5mm-rim contained two CT features with a strong Spearman correlation ( ρ ) to the tumour volume ( ρ > 0.85), the features of the signature based on the GTV 5mm-rim+3mm-ext showed only moderate correlations to the tumour volume ( ρ range: [−0.61–0.40]). This indicates that additional imaging features, which are not related to the tumour volume, may improve the risk model performance.
One limitation of this retrospective study is the different distribution of the clinical characteristics between the exploratory and validation cohort, e.g., in tumour site and UICC stage (Table 1). Despite these differences, the validation of the presented radiomic models was successful, and due to the definition of both cohorts that was based on independent clinical trials, the presented results should be more robust compared for example to a random split of the data. Furthermore, other factors related to the retrospective nature of our study may have implications on the presented results [41]. For instance, the variety of different CT imaging acquisition and reconstruction parameters may affect the feature robustness and thereby the results of risk modelling (Table A1). Therefore, open and standardised protocols for image acquisition, reconstruction, and analysis may help to increase the robustness of radiomic risk model [42,43,44]. In addition, the biological meaning of the selected imaging features within the developed signatures and the differences between the features of the tumour rim and core remains still unclear. Therefore, these open questions should be investigated systematically in the future for a better understanding of the underlying mechanisms.

5. Conclusions

In the present study, we showed that radiomic models based on the rim of locally advanced HNSCC achieved a slightly higher prognostic performance for LRC after primary radio-chemotherapy compared to models using the tumour core. This supports our initial hypothesis that the tumour rim is biologically more diverse and important treatment-related processes occur primarily in this region, which may be visible in clinical imaging data. Therefore, after additional prospective validation the consideration of tumour sub-volumes may be a promising way to improve prognostic radiomic risk models.

Author Contributions

Conceptualisation, S.L. (Stefan Leger), S.L. (Steffen Löck), M.B., M.K., C.R. and E.G.C.T.; methodology, S.L. (Stefan Leger), S.L. (Steffen Löck) and A.Z.; software, S.L. (Stefan Leger) and A.Z.; validation, S.L. (Stefan Leger), S.L. (Steffen Löck) and E.G.C.T.; investigation, S.L. (Stefan Leger); resources, S.L. (Steffen Löck); data curation, K.L., E.G.C.T., F.L., A.L., A.S., G.K., I.T., N.G., M.G., P.B., J.v.d.G., U.G., C.B., J.C.P., S.E.C., S.B. and D.Z.; writing—original draft preparation, S.L. (Stefan Leger); writing—review and editing, S.L. (Steffen Löck), A.Z., E.G.C.T., A.L., A.S., G.K., I.T., N.G., M.G., P.B., J.v.d.G., U.G., C.B., J.C.P., S.E.C., S.B., D.Z. and C.R.; visualisation, S.L. (Stefan Leger); supervision, S.L. (Steffen Löck) and E.G.C.T. All authors have read and agreed to the published version of the manuscript.

Funding

The author S.L. was supported by the Federal Ministry of Education and Research (BMBF-13GW0211D).

Conflicts of Interest

In the past 5 years, Michael Baumann attended an advisory board meeting of MERCK KGaA (Darmstadt), for which the University of Dresden received a travel grant. He further received funding for his research projects and for educational grants to the University of Dresden by Teutopharma GmbH (2011–2015), IBA (2016), Bayer AG (2016-2018), Merck KGaA (2014-open), Medipan GmbH (2014–2018). He is on the supervisory board of HI-STEM gGmbH (Heidelberg) for the German Cancer Research Center (DKFZ, Heidelberg) and also member of the supervisory body of the Charité University Hospital, Berlin. As former chair of OncoRay (Dresden) and present CEO and Scientific Chair of the German Cancer Research Center (DKFZ, Heidelberg), he has been or is still responsible for collaborations with a multitude of companies and institutions, worldwide. In this capacity, he discussed potential projects with and has signed/signs contracts for his institute(s) and for the staff for research funding and/or collaborations with industry and academia, worldwide, including but not limited to pharmaceutical corporations like Bayer, Boehringer Ingelheim, Bosch, Roche and other corporations like Siemens, IBA, Varian, Elekta, Bruker and others. In this role, he was/is further responsible for commercial technology transfer activities of his institute(s), including the DKFZ-PSMA617 related patent portfolio [WO2015055318 (A1), ANTIGEN (PSMA)] and similar IP portfolios. Baumann confirms that, to the best of his knowledge, none of the above funding sources was involved in the preparation of this paper. In the past 5 years, Krause received funding for her research projects by IBA (2016), Merck KGaA (2014–2018 for preclinical study; 2018-2020 for clinical study), Medipan GmbH (2014–2018) and by the Gert and Susanna Mayer Foundation (2019–2022). She is involved in an ongoing publicly funded (German Federal Ministry of Education and Research) project with the companies Medipan, Attomol GmbH, GA Generic Assays GmbH, Gesellschaft fü r medizinische und wissenschaftliche genetische Analysen, Lipotype GmbH and PolyAn GmbH (2019–2021). For the present manuscript, Krause confirm that none of the above mentioned funding sources were involved. In the past 5 years, Troost received funding for her research projects by Merck KGaA (2018–2020 for clinical study). She is involved in an ongoing publicly funded (German Federal Ministry of Education and Research) project with the companies Medipan, Attomol GmbH, GA Generic Assays GmbH, Gesellschaft für medizinische und wissenschaftliche genetische Analysen, Lipotype GmbH and PolyAn GmbH (2019–2021). For the present manuscript, Troost confirm that none of the above mentioned funding sources were involved. In the past 5 years, Richter and his institution OncoRay received funding from Siemens Healthineers for research collaborations and as reference center as well as lecturer. For the present manuscript, Richter confirms that none of the above mentioned funding sources were involved. D.Z., S.B.: The Depart. of Radiation Oncology Tübingen receives support through research, travel and training grants from DFG, Elekta, Philips, Siemens, Therapanacea and Sennewald. Non of these sources are relevant to the presented project. Linge is involved in an ongoing publicly funded (German Federal Ministry of Education and Research) project with the companies Medipan, Attomol GmbH, GA Generic Assays GmbH, Gesellschaft für medizinische und wissenschaftliche genetische Analysen, Lipotype GmbH and PolyAn GmbH (2019–2021). For the present manuscript, Linge confirms that this above mentioned funding source was not involved. Tinhofer served as advisor and gave lectures for Merck-Serono. For the present manuscript, Tinhofer confirms that this above mentioned funding source was not involved in the preparation of this paper.

Abbreviations

The following abbreviations are used in this manuscript:
BGLM-Coxboosted gradient linear model-Cox
BT-Coxboosted tree-Cox
C-indexconcordance index
CoxCox proportional hazard model
CTcomputed tomography
DKTK-ROGGerman Cancer Consortium Radiation Oncology Group
FDG 18 F-fluorodeoxyglucose
FMISO 18 F-fluoromisonidazole
GTVgross tumour volume
HNSCChead and neck squamous cell carcinoma
MIFSmutual information feature selection
MIMmutual information maximisation
MLAmulti-level model approach
MRImagnetic resonance imaging
MRMRminimum redundancy maximum relevance
MSR-RFmaximally selected rank statistics random forest
OSoverall survival
PETpositron emission tomography
RCTradio-chemotherapy
RFVIrandom forest variable importance
RSFrandom survival forest
UKDUniversity Hospital Dresden

Appendix A

Appendix A.1. Multi-Level Model

The multi-level model approach (MLA) was developed for assessing the differences in concordance index between the rim-based and the corresponding core-based models independent from the effects of the feature selection methods and learning algorithms, similar as in [5]. For this purpose, we defined a multi-level model consisting of three levels (L) representing the different factors:
L0 (volume):
y i , j = α method , i , j + β vol x v o l + ε vol ε vol N ( 0 , σ vol 2 ) ,
L1 (machine learning algorithms):
α method , i j = α f s , i + β learner , j + ε learner , j ε learner , j N ( 0 , σ learner , j 2 ) ,
L2 (feature selection methods):
α fs , i = β fs , i + ε fs , i ε fs , i N ( 0 , σ fs , i 2 ) .
In the multi-level model, the top level describes the effect of the tumour sub-volume. y i j is the concordance index of a bootstrap sample using feature selection method i and learning algorithm j. α method , i , j is an offset term, modelled separately in levels L1 and L2. x ( v o l ) is a contrast variable, which has the values 0 or 1 for the two specific sub-volume that are to be compared, e.g., rim versus core. β ( v o l ) is the effect of the specific sub-volume compared to the other sub-volume and has a weakly informative prior N ( 0 , 1 ) and is limited to the range [ 1 , 1 ] . The error term ε vol is modelled with a normal distribution with mean 0 and standard deviation σ vol ϵ [ 0 , 1 ] . Level L1 models the effect of learner j with feature selection method i. α fs , i is an offset modelled separately in L2. β learner , j is the effect of learner j. It has a weakly informative prior N ( 0 , 1 ) and is limited to the range [ 1 , 1 ] . The error term ε l e a r n e r , j is modelled with a normal distribution with mean 0 standard deviation σ learner , j , with σ learner , j ϵ [ 0 , 1 ] . Level L2 models the effect of feature selection method i, β fs , i . β fs , i has a weakly informative prior N ( 0.5 , 1 ) , as a concordance index takes the baseline value of 0.5 for random data, and is limited to the range [ 0 , 1 ] . The error term ε fs , i is modelled with a normal distribution with mean 0 standard deviation σ fs , i , with σ fs , i ϵ [ 0 , 1 ] . The model was fitted using Markov chain Monte Carlo in STAN 2.21.2 (Stan Development Team (2020). RStan: the R interface to Stan. R package version 2.21.1. http://mc-stan.org/), using 7 chains with 500 warm-up iterations and 500 sample iterations each. Model convergence was checked using the R-hat statistic.
Table A1. Computed tomography acquisition and reconstruction settings for the exploratory and the validation cohort.
Table A1. Computed tomography acquisition and reconstruction settings for the exploratory and the validation cohort.
Image Acquisition ParametersExploratory Cohort (n = 206)Validation Cohort (n = 85)
Voxel spacing in-plane (mm)
  0.85/0.87/0.88/0.901/2/1/10/0/0/0
  0.92/0.93/0.94/0.961/1/3/20/0/0/0
  0.97/0.98/1.17/1.27/1.363/113/21/26/290/13/0/14/58
Spacing in z-direction (mm)
  2.0/2.5/3.0/3.75/5.036/22/74/1/630/0/27/0/58
Reconstruction kernel
  B10s/B20f/s/B30f/s/B31f/s20/3/1/2/29/19/161/51/1/0/0/12/0
  B40f/s/B50s/Missing1/1/9/12/930/0/0/0/20
Mean exposure (mA)181.3 (Missing: 59)76.8 (Missing: 14)
Mean exposure time (ms)733.8 (Missing: 59)508.8 (Missing: 14)
Tube voltage (kV)
  120/130/140/Missing86/9/16/9571/0/0/14
Figure A1. Concordance indices (C-index) for the feature selection methods (columns) and the learning algorithms (rows) trained on the GTV entire , the GTV 3mm-rim , the GTV 5mm-rim and the corresponding core sub-volumes on the exploratory cohort. Furthermore, the 95% confidence intervals for each model combination are shown (in parentheses).
Figure A1. Concordance indices (C-index) for the feature selection methods (columns) and the learning algorithms (rows) trained on the GTV entire , the GTV 3mm-rim , the GTV 5mm-rim and the corresponding core sub-volumes on the exploratory cohort. Furthermore, the 95% confidence intervals for each model combination are shown (in parentheses).
Cancers 12 03047 g0a1
Figure A2. Concordance indices (C-index) for the feature selection methods (columns) and the learning algorithms (rows) trained on the GTV entire , the GTV 3mm-rim , the GTV 5mm-rim and the corresponding core sub-volumes on the validation cohort. Furthermore, the 95% confidence intervals for each model combination are shown (in parentheses).
Figure A2. Concordance indices (C-index) for the feature selection methods (columns) and the learning algorithms (rows) trained on the GTV entire , the GTV 3mm-rim , the GTV 5mm-rim and the corresponding core sub-volumes on the validation cohort. Furthermore, the 95% confidence intervals for each model combination are shown (in parentheses).
Cancers 12 03047 g0a2
Figure A3. Concordance indices (C-index) for the feature selection methods (columns) and the learning algorithms (rows) trained on the tumour extension GTV 5mm-rim+1mm-ext , GTV 5mm-rim+2mm-ext , GTV 5mm-rim+3mm-ext and GTV 5mm-rim+5mm-ext sub-volumes for the exploratory cohort. Furthermore, the 95% confidence intervals for each model combination are shown (in parentheses).
Figure A3. Concordance indices (C-index) for the feature selection methods (columns) and the learning algorithms (rows) trained on the tumour extension GTV 5mm-rim+1mm-ext , GTV 5mm-rim+2mm-ext , GTV 5mm-rim+3mm-ext and GTV 5mm-rim+5mm-ext sub-volumes for the exploratory cohort. Furthermore, the 95% confidence intervals for each model combination are shown (in parentheses).
Cancers 12 03047 g0a3
Figure A4. Concordance indices (C-index) for the feature selection methods (columns) and the learning algorithms (rows) trained on the tumour extension GTV 5mm-rim+1mm-ext , GTV 5mm-rim+2mm-ext , GTV 5mm-rim+3mm-ext and GTV 5mm-rim+5mm-ext sub-volumes for the validation cohort. Furthermore, the 95% confidence intervals for each model combination are shown (in parentheses).
Figure A4. Concordance indices (C-index) for the feature selection methods (columns) and the learning algorithms (rows) trained on the tumour extension GTV 5mm-rim+1mm-ext , GTV 5mm-rim+2mm-ext , GTV 5mm-rim+3mm-ext and GTV 5mm-rim+5mm-ext sub-volumes for the validation cohort. Furthermore, the 95% confidence intervals for each model combination are shown (in parentheses).
Cancers 12 03047 g0a4
Figure A5. Resulting p-values of the log-rank tests for loco-regional tumour control in the validation cohort for the feature selection methods (columns) and the learning algorithms (rows) trained on the GTV entire , the GTV 3mm-rim , the GTV 5mm-rim and the corresponding core sub-volumes based on median risk cut-off values using the predicted risk values. The cut-off values used for stratification were determined on the exploratory cohort and applied to the validation cohort unchanged.
Figure A5. Resulting p-values of the log-rank tests for loco-regional tumour control in the validation cohort for the feature selection methods (columns) and the learning algorithms (rows) trained on the GTV entire , the GTV 3mm-rim , the GTV 5mm-rim and the corresponding core sub-volumes based on median risk cut-off values using the predicted risk values. The cut-off values used for stratification were determined on the exploratory cohort and applied to the validation cohort unchanged.
Cancers 12 03047 g0a5
Figure A6. Resulting p-values of the log-rank tests for loco-regional tumour control on the validation cohort for the feature selection methods (columns) and the learning algorithms (rows) trained on the GTV 5mm-rim+1mm-ext , the GTV 5mm-rim+2mm , the GTV 5mm-rim+3mm-ext and the GTV 5mm-rim+5mm-ext based on median risk cut-off values using the predicted risk values. The cut-off values used for stratification were determined on the exploratory cohort and applied to the validation cohort unchanged.
Figure A6. Resulting p-values of the log-rank tests for loco-regional tumour control on the validation cohort for the feature selection methods (columns) and the learning algorithms (rows) trained on the GTV 5mm-rim+1mm-ext , the GTV 5mm-rim+2mm , the GTV 5mm-rim+3mm-ext and the GTV 5mm-rim+5mm-ext based on median risk cut-off values using the predicted risk values. The cut-off values used for stratification were determined on the exploratory cohort and applied to the validation cohort unchanged.
Cancers 12 03047 g0a6
Table A2. Configuration settings for the image feature computation and extraction.
Table A2. Configuration settings for the image feature computation and extraction.
Configuration SettingsConfiguration Value(s)
Image interpolation
  Interpolation methodCubic spline
  Voxel dimensions1.0 × 1.0 × 1.0 mm 3
  Anti-aliasing smoothing parameter β 0.98
RoI interpolation
  Interpolation methodCubic spline
  Inclusion threshold0.5
Discretisation
  Discretisation methodFixed Bin Number (FBN) of 32 bins
  Intensity Volume Histogram discretisation methodFixed Bin Number (FBN) of 1000 bins
Image transformation
  Waveletcoiflet-1
  Mean-Intensity Laplacian of Gaussiankernel widths: 1.0, 2.0, 3.0, 5.0, 6.0 mm
Texture matrices
  Grey-level Run Length Matrix (GLRLM)Calculation method: 3D
Merge method: volume merge
  Grey-level Size Zone Matrix (GLSZM)Calculation method: 3D
  Neighbourhood Grey Tone Difference Matrix (NGTDM)Calculation method: 3D
  Neighbourhood Grey Level Dependence Matrix (NGLDM)Distance for neighborhood: 1.8 voxels
Difference level: 0.0
Calculation method: 3D
  Grey Level Co-occurrence Matrix (GLCM)Distance for neighborhood: 1.0 voxels
Calculation method: 3D
Merge method: volume merge
  Grey Level Distance Zone Matrix (GLDZM)Calculation method: 3D
RoI: region of interest.
Table A3. Definition of the hyper-parameters of the feature selection methods and of the machine learning algorithms, which were used during the hyper-parameter optimisation. The hyper-parameters for the feature selection methods were kept fixed and not optimised during hyper-parameter optimisation.
Table A3. Definition of the hyper-parameters of the feature selection methods and of the machine learning algorithms, which were used during the hyper-parameter optimisation. The hyper-parameters for the feature selection methods were kept fixed and not optimised during hyper-parameter optimisation.
AlgorithmHyper-Parameter NameHyper-Parameter Value(s)
Cox proportional
hazard modelSignature size2, 3, 4, 5, 7, 10
Boosted tree and
boosted gradient modelsSignature size:2, 3, 4, 5, 7, 10
α :0.001, 0.01, 0.05
ω :lambda.min 1 , lambda.1se 2
mStop:200
Random survival forestSignature size:2, 3, 4, 5, 7, 10
ntree:2000, 5000
mtry:100
node-Size:25–50, step size 1
maxDepth:10, 15, 20, 25, 40
nSplit:1, 2, 100
splitRule:logrank, logrankscore
Maximally selected rank
statistics random forestSignature size:2, 3, 4, 5, 7, 10
ntree:2000, 5000
mtry:100
node-Size:25–50, step size 1
minprop:0.1
α :0.1, 0.5
splitRule:C, maxstat
Minimum redundancy
maximum relevance
topFeatures100
RelativeImportanceThreshold0
Mutual information
feature selection
topFeatures100
RelativeImportanceThreshold0.05
Random forest
variable importance
topFeatures20
nTree1000
K2
nRepetition50
nSteps2
nSplits1
splitrulelogrank
nodeSize45
mTry500
nVariables10
1 Minimum mean cross-validated error; 2 Error within one standard error of the minimum.
Table A4. Hyper-parameters and their optimised values for the representative models.
Table A4. Hyper-parameters and their optimised values for the representative models.
Sub-VolumeHyper-Parameter Value(s)
GTV entire
  RSF-RFVISignature size: 5
ntree: 5000
mtry: 100
node-Size: 10
maxDepth: 25
nSplit: 1
splitRule: logrankscore
GTV 5mm-rim
  RSF-MIMSignature size: 10
ntree: 5000
mtry: 100
node-Size: 5
maxDepth: 40
nSplit: 1
splitRule: logrank
GTV 5mm-core
  RSF-MIMSignature size: 5
ntree: 5000
mtry: 100
nodeSize: 7
maxDepth: 40
nSplit: 1
splitRule: logrank
GTV 5mm-rim+3mm-ext
  RSF-RFVISignature size: 7
ntree: 5000
mtry: 100
node-Size: 5
maxDepth: 15
nSplit: 1
splitRule: logrankscore
Table A5. Radiomic signatures for predicting loco-regional tumour control for the representative models based on the GTV entire , the GTV 5mm-rim and the GTV 5mm-core as well as the GTV 5mm-rim+3mm-ext tumour sub-volumes. The mathematical description and the abbreviations of features can be found in Zwanenburg et al. [25].
Table A5. Radiomic signatures for predicting loco-regional tumour control for the representative models based on the GTV entire , the GTV 5mm-rim and the GTV 5mm-core as well as the GTV 5mm-rim+3mm-ext tumour sub-volumes. The mathematical description and the abbreviations of features can be found in Zwanenburg et al. [25].
Tumour Sub-VolumeFeature NameSynonymImageCluster
GTV entire ivh_diff_v10_v90 F1 IVH wav_coif1_llhno
morph_integ_int, stat_energy F2 ¯ M/S base, wav_coif1_lllyes
ivh_v10 F3 IVH wav_coif1_llhno
rlm_gl_var_3d_avg F4 T wav_coif1_lhlno
rlm_srhge_3d_avg F5 T baseno
GTV 5mm-rim ivh_diff_v10_v90 F1 IVH wav_coif1_llhno
dzm_ldhge_3d F2 T baseno
dzm_lgze_3d, dzm_sdlge_3d, F3 ¯ T baseyes
szm_lgze_3d, szm_szlge_3d
ivh_v10 F4 IVH wav_coif1_llhno
dzm_ldhge_3d F5 T wav_coif1_lllno
morph_integ_int, stat_energy F6 ¯ M/S base, wav_coif1_lllyes
dzm_hgze_3d, szm_hgze_3d, szm_szhge_3d F7 ¯ T baseyes
cm_auto_corr_d1_3d, cm_joint_avg_d1_3d,
cm_sum_avg_d1_3d, ih_mean, F8 ¯ IH/T baseyes
ngl_hgce_d1_a0.0_3d, rlm_hgre_3d_avg
cm_info_corr2_d1_3d_avg F9 T baseno
ngt_strength_3d, rlm_rlnu_3d_avg F10 ¯ T wav_coif1_llhyes
GTV 5mm-core morph_integ_int, stat_energy F1 ¯ S base, wav_coif1_lllyes
ngt_strength_3d, rlm_rlnu_3d_avg F2 ¯ T wav_coif1_llhyes
dzm_gl_var_3d, szm_gl_var_3d F3 ¯ T wav_coif1_hhlyes
dzm_glnu_3d, szm_glnu_3d F4 ¯ T wav_coif1_llhyes
ivh_i50 F5 IVH wav_coif1_lhlno
GTV 5mm-rim+3mm-ext morph_moran_i F1 M baseno
stat_energy F2 S wav_coif1_hhlno
ih_kurt,rlm_glnu_norm_3d_avg, stat_kurt F3 ¯ IH/T wav_coif1_lhlyes
rlm_gl_var_3d_avg F4 T wav_coif1_lhhno
stat_max, stat_min, stat_range F5 ¯ S wav_coif1_hhlyes
ngl_dc_var_d1_a0.0_3d F6 T wav_coif1_lhhno
stat_cov F7 ¯ S base, wav_coif1_lllyes
l: low pass, h: high pass, wav_coif1: coiflet-1 wavelet high-/low-pass filter; M: morphological, S: first order statistic, T: texture-based feature; IH: intensity histogram, IVH: intesnity volume histogram feature; F ¯ : cluster of features represented by the mean value as a new meta-feature.

References

  1. Baumann, M.; Krause, M.; Overgaard, J.; Debus, J.; Bentzen, S.M.; Daartz, J.; Richter, C.; Zips, D.; Bortfeld, T. Radiation oncology in the era of precision medicine. Nat. Rev. Cancer 2016, 16, 234. [Google Scholar] [CrossRef] [PubMed]
  2. Aerts, H.; Velazquez, E.; Leijenaar, R.; Parmar, C.; Grossmann, P.; Carvalho, S.; Bussink, J.; Monshouwer, R.; Haibe-Kains, B.; Rietveld, D.; et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat. Commun. 2014, 5, 4006. [Google Scholar] [CrossRef]
  3. Vallières, M.; Kay-Rivest, E.; Perrin, L.; Liem, X.; Furstoss, C.; Aerts, H.; Khaouam, N.; Nguyen-Tan, P.; Wang, C.; Sultanem, K.; et al. Radiomics strategies for risk assessment of tumour failure in head-and-neck cancer. Sci. Rep. 2017, 7, 10117. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  4. Leger, S.; Zwanenburg, A.; Pilz, K.; Lohaus, F.; Linge, A.; Zöphel, K.; Kotzerke, J.; Schreiber, A.; Tinhofer, I.; Budach, C.; et al. A comparative study of machine learning methods for time-to-event survival data for radiomics risk modelling. Sci. Rep. 2017, 7, 13206. [Google Scholar] [CrossRef] [PubMed]
  5. Leger, S.; Zwanenburg, A.; Pilz, K.; Zschaeck, S.; Zöphel, K.; Kotzerke, J.; Schreiber, A.; Zips, D.; Krause, M.; Baumann, M.; et al. CT imaging during treatment improves radiomic models for patients with locally advanced head and neck cancer. Radiother. Oncol. 2019, 130, 10–17. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  6. Wu, J.; Gong, G.; Cui, Y.; Li, R. Intratumor partitioning and texture analysis of dynamic contrast-enhanced (DCE)-MRI identifies relevant tumor subregions to predict pathological response of breast cancer to neoadjuvant chemotherapy. J. Magn. Reson. Imaging 2016, 44, 1107–1115. [Google Scholar] [CrossRef]
  7. Serganova, I.; Doubrovin, M.; Vider, J.; Ponomarev, V.; Soghomonyan, S.; Beresten, T.; Ageyeva, L.; Serganov, A.; Cai, S.; Balatoni, J.; et al. Molecular imaging of temporal dynamics and spatial heterogeneity of hypoxia-inducible factor-1 signal transduction activity in tumors in living mice. Cancer Res. 2004, 64, 6101–6108. [Google Scholar] [CrossRef] [Green Version]
  8. Troost, E.G.; Laverman, P.; Philippens, M.E.; Lok, J.; van der Kogel, A.J.; Oyen, W.J.; Boerman, O.C.; Kaanders, J.H.; Bussink, J. Correlation of [18 F] FMISO autoradiography and pimonodazole immunohistochemistry in human head and neck carcinoma xenografts. Eur. J. Nucl. Med. Mol. Imaging 2008, 35, 1803–1811. [Google Scholar] [CrossRef]
  9. Zips, D.; Zöphel, K.; Abolmaali, N.; Perrin, R.; Abramyuk, A.; Haase, R.; Appold, S.; Steinbach, J.; Kotzerke, J.; Baumann, M. Exploratory prospective trial of hypoxia-specific PET imaging during radiochemotherapy in patients with locally advanced head-and-neck cancer. Radiother. Oncol. 2012, 105, 21–28. [Google Scholar] [CrossRef] [Green Version]
  10. Gatenby, R.; Grove, O.; Gillies, R. Quantitative imaging in cancer evolution and ecology. Radiology 2013, 269, 8–15. [Google Scholar] [CrossRef] [Green Version]
  11. O’Connor, J.; Rose, C.; Waterton, J.; Carano, R.; Parker, G.; Jackson, A. Imaging intratumor heterogeneity: Role in therapy response, resistance, and clinical outcome. Clin. Cancer Res. 2015, 21, 249–257. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  12. Löck, S.; Perrin, R.; Seidlitz, A.; Bandurska-Luque, A.; Zschaeck, S.; Zöphel, K.; Krause, M.; Steinbach, J.; Kotzerke, J.; Zips, D.; et al. Residual tumour hypoxia in head-and-neck cancer patients undergoing primary radiochemotherapy, final results of a prospective trial on repeat FMISO-PET imaging. Radiother. Oncol. 2017, 124, 533–540. [Google Scholar] [CrossRef]
  13. Schütze, C.; Bergmann, R.; Yaromina, A.; Hessel, F.; Kotzerke, J.; Steinbach, J.; Baumann, M.; Beuthien-Baumann, B. Effect of increase of radiation dose on local control relates to pre-treatment FDG uptake in FaDu tumours in nude mice. Radiother. Oncol. 2007, 83, 311–315. [Google Scholar] [CrossRef]
  14. Schütze, C.; Bergmann, R.; Brüchner, K.; Mosch, B.; Yaromina, A.; Zips, D.; Hessel, F.; Krause, M.; Thames, H.; Kotzerke, J.; et al. Effect of [18F] FMISO stratified dose-escalation on local control in FaDu hSCC in nude mice. Radiother. Oncol. 2014, 111, 81–87. [Google Scholar] [CrossRef]
  15. Peltanova, B.; Raudenska, M.; Masarik, M. Effect of tumor microenvironment on pathogenesis of the head and neck squamous cell carcinoma: A systematic review. Mol. Cancer 2019, 18, 63. [Google Scholar] [CrossRef] [PubMed]
  16. Alsahafi, E.; Begg, K.; Amelio, I.; Raulf, N.; Lucarelli, P.; Sauter, T.; Tavassoli, M. Clinical update on head and neck cancer: Molecular biology and ongoing challenges. Cell Death Dis. 2019, 10, 1–17. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  17. Algohary, A.; Shiradkar, R.; Pahwa, S.; Purysko, A.; Verma, S.; Moses, D.; Shnier, R.; Haynes, A.M.; Delprado, W.; Thompson, J.; et al. Combination of Peri-Tumoral and Intra-Tumoral Radiomic Features on Bi-Parametric MRI Accurately Stratifies Prostate Cancer Risk: A Multi-Site Study. Cancers 2020, 12, 2200. [Google Scholar] [CrossRef] [PubMed]
  18. Grove, O.; Berglund, A.; Schabath, M.; Aerts, H.; Dekker, A.; Wang, H.; Rios Velazquez, E.; Lambin, P.; Gu, Y.; Balagurunathan, Y.; et al. Quantitative computed tomographic descriptors associate tumor shape complexity and intratumor heterogeneity with prognosis in lung adenocarcinoma. PLoS ONE 2015, 10, 1–14. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  19. Wu, J.; Gensheimer, M.; Dong, X.; Rubin, D.; Napel, S.; Diehn, M.; Loo, B.; Li, R. Robust Intratumor Partitioning to Identify High-Risk Subregions in Lung Cancer: A Pilot Study. Int. J. Radiat. Oncol. Biol. Phys. 2016, 95, 1504–1512. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  20. Leger, S. Radiomics Risk Modelling Using Machine Learning Algorithms for Personalised Radiation Oncology. Ph.D. Thesis, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universität Dresden, 2018. [Google Scholar]
  21. Leger, S.; Zwanenburg, A.; Pilz, K.; Lohaus, F.; Linge, A.; Zöphel, K.; Kotzerke, J.; Schreiber, A.; Tinhofer, I.; Budach, C.; et al. Identification of tumour sub-volumes for improved radiomic risk modelling in locally advanced HNSCC. Radiother. Oncol. 2018, 127, 263–264. [Google Scholar] [CrossRef]
  22. Apolle, R.; Rehm, M.; Bortfeld, T.; Baumann, M.; Troost, E.G. The clinical target volume in lung, head-and-neck, and esophageal cancer: Lessons from pathological measurement and recurrence analysis. Clin. Transl. Radiat. Oncol. 2017, 3, 1–8. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  23. Linge, A.; Lohaus, F.; Löck, S.; Nowak, A.; Gudziol, V.; Valentini, C.; von Neubeck, C.; Jütz, M.; Tinhofer, I.; Budach, V.; et al. HPV status, cancer stem cell marker expression, hypoxia gene signatures and tumour volume identify good prognosis subgroups in patients with HNSCC after primary radiochemotherapy: A multicentre retrospective study of the German Cancer Consortium Radiation Oncology Group (DKTK-ROG). Radiother. Oncol. 2016, 121, 364–373. [Google Scholar] [PubMed]
  24. Shafiq-UI-Hassan, G.G.; Latifi, K.; Ullah, G.; Hunt, D.; Balagurunathan, Y.; Abdalah, M.; Matthew, B.; Goldgof, D.; Mackin, D.; Court, L.; et al. Intrinsic dependencies of CT radiomic features on voxel size and number of gray levels. Med. Phys. 2017, 44, 1050–1062. [Google Scholar] [CrossRef] [PubMed]
  25. Zwanenburg, A.; Vallières, M.; Abdalah, M.A.; Aerts, H.J.; Andrearczyk, V.; Apte, A.; Ashrafinia, S.; Bakas, S.; Beukinga, R.J.; Boellaard, R.; et al. The image biomarker standardization initiative: Standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology 2020, 295, 328–338. [Google Scholar] [CrossRef] [Green Version]
  26. Harrell, F.E., Jr.; Lee, K.L.; Mark, D.B. Multivariable prognostic models: Issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat. Med. 1996, 15, 361–387. [Google Scholar] [CrossRef]
  27. Pencina, M.J.; D’Agostino, R.B. Overall C as a measure of discrimination in survival analysis: Model specific population value and confidence interval estimation. Stat. Med. 2004, 23, 2109–2123. [Google Scholar] [CrossRef]
  28. El Naqa, I.; Grigsby, P.; Apte, A.; Kidd, E.; Donnelly, E.; Khullar, D.; Chaudhari, S.; Yang, D.; Schmitt, M.; Laforest, R.; et al. Exploring feature-based approaches in PET images for predicting cancer treatment outcomes. Pattern Recognit. 2009, 42, 1162–1171. [Google Scholar] [CrossRef] [Green Version]
  29. Dou, T.; Aerts, H.; Coroller, T.; Mak, R. Radiomic-Based Phenotyping of Tumor Core and Rim to Predict Survival in Nonsmall Cell Lung Cancer. Int. J. Radiat. Oncol. Biol. Phys. 2017, 99, S84. [Google Scholar] [CrossRef] [Green Version]
  30. Dou, T.H.; Coroller, T.P.; van Griethuysen, J.J.; Mak, R.H.; Aerts, H.J. Peritumoral radiomics features predict distant metastasis in locally advanced NSCLC. PloS ONE 2018, 13, e0206108. [Google Scholar] [CrossRef] [Green Version]
  31. Hosny, A.; Parmar, C.; Coroller, T.P.; Grossmann, P.; Zeleznik, R.; Kumar, A.; Bussink, J.; Gillies, R.J.; Mak, R.H.; Aerts, H.J. Deep learning for lung cancer prognostication: A retrospective multi-cohort radiomics study. PLoS Med. 2018, 15, e1002711. [Google Scholar] [CrossRef] [Green Version]
  32. Keek, S.; Sanduleanu, S.; Wesseling, F.; de Roest, R.; van den Brekel, M.; van der Heijden, M.; Vens, C.; Giuseppina, C.; Licitra, L.; Scheckenbach, K.; et al. Computed tomography-derived radiomic signature of head and neck squamous cell carcinoma (peri) tumoral tissue for the prediction of locoregional recurrence and distant metastasis after concurrent chemo-radiotherapy. PLoS ONE 2020, 15, e0232639. [Google Scholar]
  33. Campbell, S.; Poon, I.; Markel, D.; Vena, D.; Higgins, K.; Enepekides, D.; Rapheal, S.; Wong, J.; Allo, G.; Morgen, E.; et al. Evaluation of microscopic disease in oral tongue cancer using whole-mount histopathologic techniques: Implications for the management of head-and-neck cancers. Int. J. Radiat. Oncol. Biol. Phys. 2012, 82, 574–581. [Google Scholar] [CrossRef] [PubMed]
  34. Apolle, R.; Bijl, H.P.; Blanchard, P.; Laprie, A.; Madani, I.; Ruffier, A.; Van Elmpt, W.; Troost, E.G. Target volume delineation for adaptive treatment in HNSCC is highly variable among experts. Radiother. Oncol. 2019, 133, S655–S656. [Google Scholar] [CrossRef]
  35. Zwanenburg, A.; Leger, S.; Agolli, L.; Pilz, K.; Troost, E.G.; Richter, C.; Löck, S. Assessing robustness of radiomic features by image perturbation. Sci. Rep. 2019, 9, 1–10. [Google Scholar] [CrossRef] [Green Version]
  36. Haarburger, C.; Müller-Franzes, G.; Weninger, L.; Kuhl, C.; Truhn, D.; Merhof, D. Radiomics feature reproducibility under inter-rater variability in segmentations of CT images. Sci. Rep. 2020, 10, 1–10. [Google Scholar] [CrossRef]
  37. Pavic, M.; Bogowicz, M.; Würms, X.; Glatz, S.; Finazzi, T.; Riesterer, O.; Roesch, J.; Rudofsky, L.; Friess, M.; Veit-Haibach, P.; et al. Influence of inter-observer delineation variability on radiomics stability in different tumor sites. Acta Oncol. 2018, 57, 1070–1074. [Google Scholar] [CrossRef] [Green Version]
  38. Vaupel, P.; Kallinowski, F.; Okunieff, P. Blood flow, oxygen and nutrient supply, and metabolic microenvironment of human tumors: A review. Cancer Res. 1989, 49, 6449–6465. [Google Scholar]
  39. Ljungkvist, A.S.; Bussink, J.; Rijken, P.F.; Raleigh, J.A.; Denekamp, J.; Van Der Kogel, A.J. Changes in tumor hypoxia measured with a double hypoxic marker technique. Int. J. Radiat. Oncol. Biol. Phys. 2000, 48, 1529–1538. [Google Scholar] [CrossRef]
  40. Kanungo, T.; Mount, D.; Netanyahu, N.; Piatko, C.; Silverman, R.; Wu, A. An efficient k-means clustering algorithm: Analysis and implementation. IEEE Trans. Pattern Anal. Mach. Intell. 2002, 24, 881–892. [Google Scholar] [CrossRef]
  41. Euser, A.M.; Zoccali, C.; Jager, K.J.; Dekker, F.W. Cohort studies: Prospective versus retrospective. Nephron Clin. Pract. 2009, 113, c214–c217. [Google Scholar] [CrossRef] [Green Version]
  42. Clarke, L.P.; Nordstrom, R.J.; Zhang, H.; Tandon, P.; Zhang, Y.; Redmond, G.; Farahani, K.; Kelloff, G.; Henderson, L.; Shankar, L.; et al. The quantitative imaging network: NCI’s historical perspective and planned goals. Transl. Oncol. 2014, 7, 1. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  43. Buckler, A.J.; Bresolin, L.; Dunnick, N.R.; Sullivan, D.C.; Group. A collaborative enterprise for multi-stakeholder participation in the advancement of quantitative imaging. Radiology 2011, 258, 906–914. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  44. Buckler, A.J.; Bresolin, L.; Dunnick, N.R.; Sullivan, D.C.; Group. Quantitative imaging test approval and biomarker qualification: Interrelated but distinct activities. Radiology 2011, 259, 875–884. [Google Scholar] [CrossRef] [PubMed] [Green Version]
Figure 1. Experimental design. A multi-centre cohort of 291 patients with loco regionally advanced head and neck squamous cell carcinoma (HNSCC) was used to generate different sub-volumes based on the delineated entire tumour. In particular, the outer contour of the entire primary cross tumour volume (GTV entire ) was cropped by different margins (3 and 5 mm) to define the rim of the tumour (GTV rim ) and the corresponding core (GTV core ). Furthermore, the best performing tumour rim sub-volume was extended (GTV rim+ext ) into surrounding tissue with different distances (1, 2, 3, and 5 mm) to assess the prognostic performance of the microscopic tumour extension. The entire cohort was split into an exploratory and an independent validation cohort for risk modelling. Prognostic model performance and patient risk group stratification were assessed on the validation cohort. Selected features within the developed signatures were analysed in terms of their univariate association with loco-regional tumour control using the entire cohort.
Figure 1. Experimental design. A multi-centre cohort of 291 patients with loco regionally advanced head and neck squamous cell carcinoma (HNSCC) was used to generate different sub-volumes based on the delineated entire tumour. In particular, the outer contour of the entire primary cross tumour volume (GTV entire ) was cropped by different margins (3 and 5 mm) to define the rim of the tumour (GTV rim ) and the corresponding core (GTV core ). Furthermore, the best performing tumour rim sub-volume was extended (GTV rim+ext ) into surrounding tissue with different distances (1, 2, 3, and 5 mm) to assess the prognostic performance of the microscopic tumour extension. The entire cohort was split into an exploratory and an independent validation cohort for risk modelling. Prognostic model performance and patient risk group stratification were assessed on the validation cohort. Selected features within the developed signatures were analysed in terms of their univariate association with loco-regional tumour control using the entire cohort.
Cancers 12 03047 g001
Figure 2. Kaplan–Meier curves for the prediction of loco-regional tumour control of the representative models based on the (a) entire primary gross tumour volume (GTV entire ); (b) 5 mm rim of the tumour (GTV 5mm-rim ); (c) corresponding tumour core (GTV 5mm-core ) and (d) 3 mm extension of the 5 mm tumour rim (GTV 5mm-rim+3mm-ext ) sub-volumes for patients of the validation cohort. Patients were stratified into low (LR) and high (HR) risk groups based on the median risk of loco-regional recurrence determined on the exploratory cohort.
Figure 2. Kaplan–Meier curves for the prediction of loco-regional tumour control of the representative models based on the (a) entire primary gross tumour volume (GTV entire ); (b) 5 mm rim of the tumour (GTV 5mm-rim ); (c) corresponding tumour core (GTV 5mm-core ) and (d) 3 mm extension of the 5 mm tumour rim (GTV 5mm-rim+3mm-ext ) sub-volumes for patients of the validation cohort. Patients were stratified into low (LR) and high (HR) risk groups based on the median risk of loco-regional recurrence determined on the exploratory cohort.
Cancers 12 03047 g002
Figure 3. Heatmaps showing different expression patterns of the radiomic features of the developed signatures for the representative models based on the (a) entire primary gross tumour volume (GTV entire ); (b) 5 mm rim of the tumour (GTV 5mm-rim ); (c) corresponding tumour core (GTV 5mm-core ); and (d) 3 mm extension of the 5 mm tumour rim (GTV 5mm-rim+3mm-ext ) sub-volumes. Feature expression values are sorted according to the predicted risk and the risk group based on the determined median risk cut-off values. Loco-regional tumour control (LRC) during follow-up (yes, light; no, dark) and features with a significant association with LRC are shown (* p < 0.05 and ** p < 0.001). A detailed description of the feature abbreviations can be found in Table A5. Abbreviations: F ¯ cluster feature consisting of several features represented by the mean value as a new meta-feature, F S first order statistical feature, F M morphological, F IH intensity histogram, F IVH intensity volume histogram and F T texture feature.
Figure 3. Heatmaps showing different expression patterns of the radiomic features of the developed signatures for the representative models based on the (a) entire primary gross tumour volume (GTV entire ); (b) 5 mm rim of the tumour (GTV 5mm-rim ); (c) corresponding tumour core (GTV 5mm-core ); and (d) 3 mm extension of the 5 mm tumour rim (GTV 5mm-rim+3mm-ext ) sub-volumes. Feature expression values are sorted according to the predicted risk and the risk group based on the determined median risk cut-off values. Loco-regional tumour control (LRC) during follow-up (yes, light; no, dark) and features with a significant association with LRC are shown (* p < 0.05 and ** p < 0.001). A detailed description of the feature abbreviations can be found in Table A5. Abbreviations: F ¯ cluster feature consisting of several features represented by the mean value as a new meta-feature, F S first order statistical feature, F M morphological, F IH intensity histogram, F IVH intensity volume histogram and F T texture feature.
Cancers 12 03047 g003
Table 1. Patient characteristics of the exploratory and the independent validation cohort.
Table 1. Patient characteristics of the exploratory and the independent validation cohort.
Clinical VariableExploratory CohortValidation Cohortp-Value
Number of patients20685-
Gender
  male174740.70 2
  female3211
Age in years
  median59.055.00.023 3
  range39.2–84.537.0–76.0-
cTN staging
  T stage 1/2/3/42/23/51/1302/9/30/440.21 1
  N stage 0/1/2/3/missing30/7/154/15/09/8/64/3/10.097 1
UICC stage 2010
  I/II/III/IV0/0/15/1911/2/9/730.039 1
GTV (cm 3 )
  median29.140.60.067 3
  range4.5–321.72.7–239.1-
Tumour site
  oropharynx/oral cavity/
  hypopharynx/larynx93/51/62/029/23/28/50.003 3
p16 status
  negative/positive/missing148/28/3052/5/280.26 1
Loco-regional tumour recurrence84 (41%)28 (33%)0.26 3
Follow up time (months)
  median21.224.3-
  range1.2–131.91.3–107.20.64 3
Abbreviations: T, clinical tumour stage; N, clinical nodal stage; UICC, Union internationale contre le cancer; Gy, Gray; DNA, deoxyribonucleic acid; GTV, primary gross tumour volume; 1 χ 2 test; 2 exact Fisher test; 3 Wilcoxon–Mann–Whitney test.
Table 2. Median concordance indices (C-index) of radiomic models using the entire gross tumour volume (GTV entire ) and the different sub-volumes, i.e., tumour rim (GTV rim ), extended tumour rim (GTV rim+ext ) and tumour core (GTV core ) for the endpoint loco-regional tumour control. Results are presented for the exploratory and the validation cohort. Median results over all feature selection methods and learning algorithms are shown (top) as well as C-indices of the representative model combinations and the p-values of the log-rank tests of stratified patient groups (bottom).
Table 2. Median concordance indices (C-index) of radiomic models using the entire gross tumour volume (GTV entire ) and the different sub-volumes, i.e., tumour rim (GTV rim ), extended tumour rim (GTV rim+ext ) and tumour core (GTV core ) for the endpoint loco-regional tumour control. Results are presented for the exploratory and the validation cohort. Median results over all feature selection methods and learning algorithms are shown (top) as well as C-indices of the representative model combinations and the p-values of the log-rank tests of stratified patient groups (bottom).
Tumour Sub-VolumeExploratory CohortValidation Cohort
All All GTV 20 cm 3 GTV > 20 cm 3
(n = 206) (n = 85)(n = 20)(n = 65)
GTV entire 0.75 ± 0.05 0.61 ± 0.040.61 ± 0.070.59 ± 0.02
GTV 3mm-rim 0.76 ± 0.06 0.63 ± 0.030.62 ± 1.000.61 ± 0.02
GTV 3mm-core 0.74 ± 0.06 0.60 ± 0.020.63 ± 0.050.57 ± 0.02
GTV 5mm-rim 0.76 ± 0.06 0.65 ± 0.020.67 ± 0.070.61 ± 0.01
GTV 5mm-core 0.72 ± 0.04 0.59 ± 0.010.69 ± 0.070.57 ± 0.04
GTV 5mm-rim+1mm-ext 0.76 ± 0.06 0.62 ± 0.030.58 ± 0.070.65 ± 0.03
GTV 5mm-rim+2mm-ext 0.76 ± 0.07 0.63 ± 0.030.67 ± 0.040.66 ± 0.04
GTV 5mm-rim+3mm-ext 0.75 ± 0.08 0.64 ± 0.050.61 ± 0.050.65 ± 0.05
GTV 5mm-rim+5mm-ext 0.75 ± 0.07 0.63 ± 0.040.65 ± 0.060.62 ± 0.05
GTV: primary gross tumour volume; sd: standard deviation; CI: confidence interval.
Table 3. Concordance indices (C-index) of the representative radiomic combinations and the p-values of the log-rank tests of stratified patient based on the entire gross tumour volume (GTV entire ) and the different sub-volumes, i.e., tumour rim (GTV rim ), extended tumour rim (GTV rim+ext ) and tumour core (GTV core ) for the endpoint loco-regional tumour control. Results are presented for the exploratory and the validation cohort.
Table 3. Concordance indices (C-index) of the representative radiomic combinations and the p-values of the log-rank tests of stratified patient based on the entire gross tumour volume (GTV entire ) and the different sub-volumes, i.e., tumour rim (GTV rim ), extended tumour rim (GTV rim+ext ) and tumour core (GTV core ) for the endpoint loco-regional tumour control. Results are presented for the exploratory and the validation cohort.
Representative ModelExploratory CohortValidation Cohort
(n = 206) (n = 85)
C-Index95% CIp-ValueC-Index95% CIp-Value
GTV entire
  RSF-RFVI0.75[0.71–0.81]<0.0010.63[0.49–0.67]0.012
GTV 3mm-rim
  BT-Cox-MRMR0.76[0.71–0.82]<0.0010.63[0.52–0.70]0.005
GTV 3mm-core
  BT-Cox-MIM0.75[0.70–0.80]<0.0010.63[0.50–0.70]0.069
GTV 5mm-rim
  RSF-MIM0.77[0.72–0.82]<0.0010.66[0.52–0.69]0.006
GTV 5mm-core
  RSF-MIM0.71[0.66–0.77]<0.0010.61[0.49–0.69]0.11
GTV 5mm-rim+3mm-ext
  RSF-RFVI0.77[0.69–0.80]<0.0010.67[0.60–0.77]<0.001
GTV: primary gross tumour volume; sd: standard deviation; CI: confidence interval; RSF: random survival forest; BT-Cox: boosted tree-Cox proportional hazard model; RFVI: random forest variable importance; MRMR: minimum redundancy maximum relevance; MIM: mutual information maximisation.
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Leger, S.; Zwanenburg, A.; Leger, K.; Lohaus, F.; Linge, A.; Schreiber, A.; Kalinauskaite, G.; Tinhofer, I.; Guberina, N.; Guberina, M.; et al. Comprehensive Analysis of Tumour Sub-Volumes for Radiomic Risk Modelling in Locally Advanced HNSCC. Cancers 2020, 12, 3047. https://0-doi-org.brum.beds.ac.uk/10.3390/cancers12103047

AMA Style

Leger S, Zwanenburg A, Leger K, Lohaus F, Linge A, Schreiber A, Kalinauskaite G, Tinhofer I, Guberina N, Guberina M, et al. Comprehensive Analysis of Tumour Sub-Volumes for Radiomic Risk Modelling in Locally Advanced HNSCC. Cancers. 2020; 12(10):3047. https://0-doi-org.brum.beds.ac.uk/10.3390/cancers12103047

Chicago/Turabian Style

Leger, Stefan, Alex Zwanenburg, Karoline Leger, Fabian Lohaus, Annett Linge, Andreas Schreiber, Goda Kalinauskaite, Inge Tinhofer, Nika Guberina, Maja Guberina, and et al. 2020. "Comprehensive Analysis of Tumour Sub-Volumes for Radiomic Risk Modelling in Locally Advanced HNSCC" Cancers 12, no. 10: 3047. https://0-doi-org.brum.beds.ac.uk/10.3390/cancers12103047

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop