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
) [
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
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
. In a further study, Wu et al. [
19] identified different tumour sub-volumes using computed tomography (CT) and
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
, 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
, 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
to account for potential sub-microscopic spread, leading to the clinical target volume [
22].
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, 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
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
may affect radiomic features and in turn the results of radiomic analyses. In the current study, the GTV
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
and the GTV
(C-index: 0.65 and 0.67, respectively). While the radiomic signature based on the GTV
contained two CT features with a strong Spearman correlation (
) to the tumour volume (
> 0.85), the features of the signature based on the GTV
showed only moderate correlations to the tumour volume (
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.