Key words
Radiomics - biomarkers - head and neck cancer
1. Introduction
Precision medicine aims at most accurately defining diseases in order to find personalized
and individual therapies. This approach should improve healing chances and/or lead
to a reduced spectrum of side effects. Diseases that in earlier times had been defined
clinically are more precisely diagnosed by histological, molecular, or a broad spectrum
of clinical biomarkers. In this way, biomarkers and increasingly also biomarker signatures
consisting of a typical pattern of different characteristics influence more and more
the clinical routine. Their prognostic and diagnostic relevance, however, may vary
significantly with regard to their specificity and sensitivity. An influence is already
observed with the definition of the patient cohort, the analytic method and its variance
as well as the definition of the respective correlation. The degree of standardization
has another effect on all levels: the method of data collection, the type of data
acquisition, data processing as well as evaluation. Broad screening methods producing
enormous amounts of data are increasingly applied. They allow simultaneous measurement
of many parameters and their comparison with regard to the applicability as biomarkers
and thus open the option to define various biomarker combinations for different patient
groups.
Hereby vast amounts of data arise that necessitate providing large storage capacities
and developing suitable software for valid analysis. It is also beneficial to have
a safe exchange of those data on an interdisciplinary and trans-regional level. The
extraction of parameters that are decisive for the respective question can only be
performed by IT-based data analysis, statistics, and modeling. The broad data collection
bears the possibility to analyze the data pool with regard to different issues and
in different directions. However, large data pools with sufficiently large patient
cohorts are necessary for defining significant biomarkers. The data analysis may be
a relevant challenge that is more complex than the analysis of the material itself
and that is also associated with multiple potential sources of errors.
In the last years, those screening examinations were included increasingly in the
biomarker analysis under the umbrella term of “omics”. According to the definition,
the overall primary analysis is relatively unselective. The suffix “omics” characterizes
different sources of large data volumes whereby this suffix is preceded by naming
the original materials of data collection for definition. This leads for example to
the description of the “genomics” in the context of genetic expression analysis or
to “transcriptomics” for investigations on the RNA level; protein analyses become
“proteomics” or also “metabolomics” when the metabolome is analyzed. In analogy, comparably
newly defined “radiomics” find their way into clinical research and increasingly even
into clinical routine.
Defined characteristics of a disease contain information about prognosis and diagnosis
with regard to the status, the outcome, and the therapeutic response. Those variables
are analyzed in the context with clinical data and the course of the disease while
the clinical data still serve as diagnostic and prognostic parameters.
Valid biomarkers should be easily accessible, measurable, and reproducible. So their
stability with regard to the measurements should be verified. The definition should
be performed in a representative, possibly large and standardized cohort where well-defined
parameters are sufficiently assessed. After definition of the biomarker signature,
it should be reevaluated in a second cohort – preferably independently in a second
institution. Further prospective validation is reasonable in order to confirm the
reliability.
Biomarkers are an essential prerequisite for personalized therapy. For some diseases
(e. g. breast cancer, prostate cancer) it is possible to define single biomarkers
with high significance but because of the complexity of pathogenesis, also in this
context increasingly biomarker matrices are applied.
These may originate from one or several data sources. Clinical, genomic, histopathological,
and other markers may be combined. These combinations are analyzed ideally by means
of software that allows quantification, configuration, and visualization. Thus, large
data volumes provide various options for the definition of appropriate markers or
marker patterns in different stages of the disease, on different steps of the analysis,
and of most diverse materials and data sources. Unfortunately, however, also the same
amount of possible misinterpretations, analysis bias, methodical and statistical sources
of errors must be considered that are difficult to identify and discover due to their
high complexity.
Up to now, it is mostly necessary for analyses to gain material from solid tumors
– e. g. tissue biopsies. These specimens undergo genomics, proteomics, or metabolomics
with the help of a broad spectrum of analytic methods (e. g. next generation sequencing,
NGS) and thousands of data are collected. Radiomics as a relative newcomer on the
scene of biomarkers have the great advantage that no invasive specimen gaining is
required.
Using mathematical algorithms, a quantitative high-throughput extraction of radiological
features based on meta-datasets (DICOM format) is performed. Often these image features
cannot be perceived with the human eye and so they can only be assessed in an IT-supported
way [1]. Historically, radiomics originate from computer-assisted diagnosis and detection
systems (CAD) of the 1980s and 1990s [2]
[3]. The difference, however, is the extracted data volume and the type of combination
with clinical, histological, or genomic data volumes. While CAD systems only provide
answers to single questions regarding diagnosis and detection of a disease, radiomics
allow generation of large data volumes from imaging such as computed tomography (CT),
magnetic resonance imaging (MRI), or positron emission tomography (PET). Nonetheless,
there are older investigations that already meet the requirements of radiomics to
a certain extent without using that term at the time of examination and thus they
had not been defined as such. A PubMed research provides the first results for the
key word of “radiomics” in 2012. The research group of Lambin et al. published an
article entitled “Radiomics: extracting more information from medical images using
advanced feature analysis” [4]. This title contains the exact definition of this new term.
1.1 Principles of radiomics
Fortunately, the typically applied imaging technique is performed routinely. It is
generally accessible and ideally already used in the context of diagnosis. However,
the basic data of imaging have to be available comprehensively for processing and
furthermore imaging has to be performed digitally according to current standards with
adequate accuracy and ideally without artefacts ([Fig. 1]). In order to achieve this objective, innovations and increasing standardization
of medical imaging have contributed to make those new methods possible. Only in this
way, sufficiently large data pools are available in single institutions to establish
radiomic signatures. In addition, modern hardware, the use of comparable radiocontrast
media, and the standardization of imaging protocols are important factors to enable
quantitative analysis and to apply specific software to this end.
Fig. 1 Schematic, simplified workflow for creation of a radiomic signature.
Standardized imaging protocols and the routine application of modern software are
essential for reproducible biomarker signatures gained from radiomics that can be
compared on a multicenter level. Only in this way, the broad quantitative analysis
becomes possible [4]. Hence, imaging performed in clinical routine may be used as gigantic source for
data analysis. In order to assess the volume of the potentially available data, one
must understand that each of those data sources – regardless of 2- or 3-dimensional
– of each patient contains millions of voxels and hundreds of features that are available
for radiomics analyses [5]. The data are ready for analysis, exist and are automatically newly generated every
day. Additional material collection such as biopsies and potentially expensive tissue
analyses are not needed.
Radiomics data can be extended by using a combination of different imaging procedures
as for example CT/PET, “dual source/dual energy” CT scans as well as the application
of radiological markers. Hereby the tissue itself, disease-specific markers as well
as increasingly also biological processes are superiorly visualized. This allows generating
different and additional characteristics from the obtained images.
Another variation of radiomics is radiogenomics. Radiomics are based on the interesting
hypothesis that cellular and phenotypic tissue properties correspond to specific radiomic
features and are displayed in imaging because radiological images are nothing else
than tissue depictions [4]. The more differentiated and the finer the examination methods are, the more accurate
and specific is the possible depiction and the more findings are provided. Tissue
specification becomes more and more accurate with regard to macroscopic, microscopic-histological,
immunohistological, electron microscopic, and molecular aspects ([Fig. 2]). The principle of radiogenomics is a logic consequence of this idea and is based
on the hypothesis that even proteogenomic cell and tissue characteristics are – directly
or indirectly – visualized by imaging procedures. The assumption in this context is
not that single mutations may be visualized e. g. in CT scans but that tissue characteristics
are induced by certain proteogenomic constellations. For example, increased regulation
of cell cycle genes might trigger a heterogenic tissue structure. The idea is certainly
fascinating that proteogenomic and cellular characteristics and thus also local and
individual differences have equivalents in imaging and that basic data may provide
currently unknown additional information, which can be assessed with specific algorithms,
i. e. “methods”, and adequately correlated and set into a context. Nonetheless it
must be stated that also a microscopic image without additional information such as
e. g. immunohistochemistry, can provide only limited findings. Since even imaging
is visualization of tissue, the information is limited.
Fig. 2 a Description of a cervical lymph node metastasis after importing the imaging into
the segmentation software (courtesy of Prof. S. Wesarg, Fraunhofer Institute Darmstadt,
Germany).
Fig. 2 b Description of a cervical lymph node metastasis after semi-automatic segmentation
(in red).
Some studies could already reveal that radiomic analyses were able to extract and
correlate equivalents to cellular, genetic, or phenotypic characteristic from classic
imaging procedures [1]. An additional challenge is the fact that radiogenomics combine 2 “omics”, i. e.
radiomics and genomics. This leads to a very large data pool and its sufficient evaluation
and application can only be performed in a professional and IT-based way. Separately
considered, the genomics data as well as the radiomics data have various advantages
and their combination may generate useful additional information. Via genomics it
is possible to identify a very detailed genetic pattern and thus an extract of molecular
processes on the cellular level – broken down to DNA or RNA. However, this is limited
to specific, rather small areas or cell types and tissue or tumor parts.
From these analyses specific biomarkers relevant for the discrete question have to
be defined. Already this objective is a great challenge. It is certainly not realistic
to aim at one-to-one correlation of genomic examination with radiomics. But if it
is successful to adequately correlate the genomic data that are relevant for a problem
and to perform genetic/molecular subtyping on a radiological level, this alone would
be an enormous benefit. With radiogenomics, for example a tumor could be holistically
characterized because in the radiological image not only single areas but the entire
affected tissue would be assessed. A combination of both methods would then provide
significantly more information also on a molecular biological level. If radiogenomics
retrieved important molecular phenomena in equivalence to laboratory diagnostic methods,
even avoiding molecular tissue diagnostics might be possible. This could finally lead
to the potential avoidance of invasive biopsies, lower costs, and less staff- and
material-related efforts as well as improve the patient satisfaction.
Currently radiomics are mainly applied in oncology for alternative characterization
of solid carcinomas. However, they have the potential to serve as biomarkers for benign
diseases with one or several classifiable correlates in imaging – e. g. in the context
of Menière’s disease [6] or functional disorders of the parotid gland after radiation [7].
In the following paragraphs, examples of some radiomics applications in the field
of oncology will be focused and the state-of-the-art in head and neck oncology will
be described. For better understanding of the methods and their problems, the radiomics
workflow will further be described and its potential sources of variation and errors
will be discussed.
2. Radiomics and Tumors
Especially in oncology, radiomics are very well received. In this context, the outcome,
histology, subtyping, or even therapy response are correlated with imaging features.
There are already older investigations that date back to the 1970s and that follow
the same principle – although this was not called “radiomics” at that time. The distinction
of the terms is certainly vague. However, because of the possibilities of data storage
and processing that were not available to the current extent, these articles are significantly
more limited regarding their spectrum of analysis. Imaging techniques that were used
included CT scan, MRI, PET/CT but also conventional X-ray, mammography, or ultrasound.
The number of publications is manageable although if has greatly increased in the
last years with definition of the term of radiomics and an associated workflow as
well as forward-looking expectations. A high variation is found in the size of the
patient cohorts and the study design. Generally speaking, older studies deal with
smaller cohorts than more recent ones – finally also due to capacities and possibilities
of data processing and storage. Although most trials are based on retrospective datasets,
a tendency is observed that radiological signatures are validated in second or even
prospective datasets – which is highly desirable.
It is remarkable that specific tumor entities are significantly more present in this
field than others. There are comparably many publications for example on lung and
breast cancer, whereas other also frequently occurring entities such as cervix cancer
or lymphomas are rather underrepresented. Therefore, some articles about well-investigated
tumors such as lung and breast cancer will be elucidated here and other less dominant
entities will not be described in this context.
2.1 Lung cancer
Already in the 1970ies, before the era of “omics”, Sutton and Hall correlated structural
analyses in radiography of the lung with different pathologies. Their objective was
to evaluate the feasibility of automatized diagnostics of chest radiography. However,
in comparison to current possibilities, the dataset was quite restricted because IT-based
analysis as we use it today was not possible. Thus, the article may be considered
rather as a precursor than as an example for modern radiomics [8].
In 2008, Al-Kadi and Watson differentiated aggressive lung carcinomas from non-aggressive
ones based on CT scans by means of structure-based characteristics in 15 patients
[9].
Since 2010, the research team of Ganeshan et al. is continuously working on radiomics
of lung carcinomas. In 2010, they published a pilot study that encompassed 18 patients
with non-small cell lung cancer. Statistically processed imaging properties (middle
greyscale, entropy, uniformity) could be correlated with the tumor stage and its glucose
metabolism [10]. In a follow-up trial, the tumor uniformity could be correlated with the survival
of 54 patients. Further investigations revealed that the structural properties could
also be matched with histopathological characteristics beside the clinical ones [11]. By applying additional statistical evaluation methods, further histopathological
properties could be correlated to the structural analysis of the CT scans [12].
Other publications concerning non-small cell lung cancer were provided by Aerts et
al. [13]. They investigated 440 radiomic features in non-small cell lung and head and neck
cancer. That aspect will be elucidated later in this article when focusing on head
and neck cancer. The research team could establish a predictive signature in lung
cancer for the survival, the histology, and the tumor stage. Recently, those parameters
were confirmed for the survival in another patient cohort and the transferability
on the modalities of planning-CT and CBCT (cone beam computed tomography) verified.
This implies that radiomic signatures are potentially applicable in different modalities
which may be of high clinical value [14].
Based on a cohort of 182 patients with adenocarcinomas of the lung, the same group
could show that a radiomic signature with 33 markers may predict metastatic spread,
and another signature encompassing 12 features the survival [15]. The analysis of other characteristics of the complexity of tumor type and heterogeneity
could be correlated with the overall survival in a patient cohort and confirmed in
an additional one [16].
Also in pulmonary adenocarcinomas (n=431), Yuan et al. compared 20 selected radiomic
biomarkers in CT scans with volumetric analysis in order to differentiate distinct
phenotypes (carcinoma in situ versus minimally invasive carcinoma versus invasive
carcinoma). In this context, the radiomic signature (accuracy: 80.5%) was superior
to volumetric analysis (accuracy: 69.5%) [17]. The comparison of these methods was thus decided in favor of radiomics.
Zhang et al. optimized the radiomic signature for the prediction of recurrences, death,
and recurrence-free survival in non-small cell lung cancer by varying different methods
for parameter selection and classification [18]. In this way, they could show that the applied statistical methods have a significant
impact on the definition and relevance of radiological biomarkers. In addition, the
modalities within one imaging variant are crucial. Different data quality and relevance
regarding the prognosis of recurrences of lung cancer were revealed by Huynh et al.
comparing different CT modalities (static versus respiration-adapted) [19].
Even radiogenomics have already been investigated in lung cancer. Aerts et al. could
find a high correlation of genomic data obtained from gene-set enrichment analysis
(GSEA) with radiomic parameters in patients with non-small cell lung cancer. Two characteristics
of radiomic heterogeneity could be correlated with cell cycle genes that lead to the
development of heterogeneous tumors and increased proliferation [13]. This substantiates the hypothesis that proteogenomic phenomena can be displayed
directly and indirectly in imaging data.
Only recently, the same group elucidated in pulmonary adenocarcinomas that a CT-based
radiomic signature (heterogeneity-based) in 353 patients could predict the EGFR (epidermal
growth factor receptor) status. This signature was validated in a second cohort of
352 patients. A combination with a clinical data model further improved the accuracy.
A signature intended to differentiate KRAS-positive from KRAS-negative tumors in the
same cohorts, was also significant but with a clearly poorer accuracy than the EGFR-associated
signature [20]. In diffusion-weighted MRI, Yuan et al. could confirm the EGFR mutation status of
pulmonary adenocarcinomas [21].
Not only outcome parameters and biological tissue typing may be displayed by radiomics.
Tools to support decision-making in therapy planning would be of great clinical value.
In cases of non-small cell lung cancer, a correlation of the response to radiotherapy
or radiochemotherapy with the overall survival could be found in PET and PET-CT by
means of radiomic biomarkers [22]
[23]. An additional benefit and even sometimes better performance compared to imaging
properties of the primary tumor (n=85) resulted from an analysis of imaging parameters
of lymph node metastases (n=178) of stage II-III non-small cell lung cancer for the
prediction of the response to neoadjuvant radiochemotherapy [24].
The versatility of imaging-based biomarkers is very well reflected in radiomics-based
studies on lung cancer. Overall, CT-based and increasingly also PET-CT trials dominate
the investigations of lung cancer. The results are very promising and encompass the
differentiation of malignant and benign lesions, the elucidation of genetic and histological
foundations as well as clinically oriented prognoses of outcome and therapy response.
2.2 Breast cancer
Already early, breast cancer has been evaluated with regard to the significance of
imaging because mammography has been used for screening for several decades. A rapid
and precise differentiation of benign and malignant lesions based on structural characteristics
as endpoints as well as the possibility of implementing automated screening have been
evaluated for a long time. Structural analyses of mammograms reach back to the 1980ies
and were continued in the 1990ies and 2000ies [25]
[26]
[27]
[28]
[29]
[30]
[31]. Since the 1990ies, structural analysis has also been successfully implemented in
ultrasound diagnostics for differentiation of breast cancer [32]
[33]
[34]. Already in 1993, Garra et al. achieved a sensitivity of 100% and a specificity
of 80% regarding the detection of malignant lesions in their examinations of 80 patients
[32]. A recent investigation could analyze 364 structural parameters by means of sonoelastography
in 42 patients suffering from breast cancer and 75 patients with benign lesions. Seven
sonoelastic characteristics were selected that could predict malignancy with a sensitivity
of 85.7% and a specificity of 89.3% [35]. A fluent transition of early structural analyses to modern radiomics can be observed.
In contrast to the detection of lung cancer where computed tomography plays a major
role, magnetic resonance imaging is the adequate procedure for soft tissue visualization
in breast cancer. Also here, the first structural analyses date back to the 1990ies.
Already in 1997, Sinha et al. could differentiate benign from malignant breast lesions
based on 8 structural characteristics combined with the patients' age with a sensitivity
of 93% and a specificity of 95% [36]. Also in the following studies that had been performed with the purpose of differentiating
malignant breast tumors, structural analyses could achieve good results. However,
those were retrospective data analyses with a limited number of patients [37]
[38]
[39]. Cai et al. initiated a study with a relatively large cohort of 234 patients in
which they could differentiate breast cancer from benign lesions with a sensitivity
of 85% and a specificity of 89%. By means of 3 classic machine-learning algorithms,
28 structural parameters were examined and in order to avoid redundancy and to achieve
improved significance, they were reduced to 5 features. The established 5 structural
parameters in the diffusion-weighted MRI (apparent diffusion coefficient, sum average,
entropy, elongation, sum variance) were validated in a second cohort of 93 patients
with a sensitivity of 69% and a specificity of 91% [40]. Recently, Bickelhaupt et al. could show that one specific radiomic signature usefully
completed the analytic significance of the apparent diffusion coefficient alone for
differentiation of malignant lesions in MRI [41]. Holli et al. succeeded in finding a histological subtyping between lobular and
ductal breast cancer, however, only within a pilot study of 20 patients (n=10 suffering
from ductal breast cancer, n=10 with lobular breast cancer) [42].
In terms of radiogenomics, a correlation of MRI-based characteristics regarding the
subtyping of 91 biopsies of invasive breast carcinomas with genomic data (TCGA/TCIA:
The Cancer Genome Atlas/The Cancer Imaging Archive) in a multicenter analysis of the
National Cancer Institute could be found. By means of radiomics, Wang et al. identified
breast cancers in DCE-MRI that did not have the typical genomic markers (“triple negative”)
[43]. A combined approach of 38 radiomic parameters and 144 genetic properties was chosen
by Guo et al. that were tested in combination and against each other. Radiomic features
were more suitable for predicting the tumor stage whereas genomic features better
described the receptor status. The data of the 91 included patients originated from
the TCIA and TCGA databases. However, the research team admitted a reduced significance
of their trial because of the limited number of patients [44]. Recently, also Li et al. tested the predictive significance of MRI-based radiomics
against genetic tests applied in clinical routine for breast cancer (MammaPrint, Oncotype
DX, PAM50 Gene Assay) based on data of 84 patients and came to the conclusion that
the radiomics-based testing might play a role in the prognosis of recurrences [45].
The response of breast cancer to chemotherapy was evaluated by several research groups.
Ahmed et al. and Parikh et al. could find significant differences between chemotherapy
responders and failures based on 8 and 2 (entropy and uniformity) MRI-based structural
parameters, respectively [46]
[47]. A recent trial of Braman et al. did not only rely on tumor-based radiomics for
prediction of therapy response, but also examined the tumor-surrounding tissue. Insufficient
response to neoadjuvant chemotherapy was associated with a higher peritumoral heterogeneity.
The combined examination could significantly predict the treatment response, independently
from the receptor status [48]. In this way, the field that is considered in radiomic examinations was extended
by this study. Not only the tumor itself provides relevant data.
In summary, it can be stated that the variance of examination methods for breast cancer
of which radiological structural parameters can be obtained is higher than for lung
cancer. Since it is a primarily soft tissue-associated tumor, the classic examination
methods of mammography, ultrasound including sonoelastography, and MRI are first-line
techniques. Also in this context, studies have already been conducted that differentiate
not only benign and malignant lesions as the endpoint, but that rather emphasize histological
and genetic foundations and predict the clinical outcome as well as the therapy response.
Data sources that were used for several analyses were not only own imaging data but
larger accessible databases such as TCIA and TCGA. In order to obtain statistically
more reliable results in the future and to validate radiomic signatures also trans-regionally,
they are certainly – even for other tumor entities – a suitable data source that should
be taken into consideration.
Fewer radiomics trials exist for other solid carcinomas such as cervix carcinomas
[49], liver carcinomas [50], colon carcinomas [51], and prostate carcinomas [52]
[53]
[54]
[55]. Concerning glioblastomas and gliomas, radiomics-based correlations could be found
with molecular information such as the EGFR status and the isocitrate dehydrogenase
1 (IDH1) status [56]
[57]
[58]. Radiological structural characteristics could also be determined for renal cell
carcinomas that correlated with the mutation status of BAP1 (BRCA2-associated protein
1) gene, VHL (von Hippel-Lindau) gene, or KDM5C gene as well as EGFR receptor status
[59]
[60]
[61]. In a proof-of-concept pilot study that analyzed the structural characteristics
of FLT-PET/MRI of patients with metastatic renal cell carcinoma, the therapeutic response
to the receptor tyrosine kinase inhibitor Sunitinib could be predicted [62].
2.3 Radiomics in head and neck cancer
Regarding the head and neck area, radiomics-based investigations already exist for
esophageal cancer, nasopharyngeal cancer, and “classic” squamous cell carcinomas of
the oro- and hypopharynx, larynx, and the oral cavity.
In a cohort of 41 patients suffering from esophageal cancer, Tixier et al. evaluated
the therapy response to combined radiochemotherapy (5-fluorouracil with carbo- or
cisplatin). They analyzed 38 radiomic parameters of pretherapeutically performed whole-body
(18)F-FDG PET examinations. Hereby, complete and partial therapy responders as well
as failures could be identified more reliably than with standard uptake values (SUV)
alone [63].
The research team of Zhang et al. recently published 2 articles on radiomics of nasopharyngeal
carcinomas. They were based on MRI for which 870 radiomic features were evaluated
per patient. The first study encompassed 110 patients and 6 methods for parameter
selection and 9 classification methods were analyzed. An optimal machine-learning
method was identified in order to perform biomarker screening of nasopharyngeal carcinomas
[64]. In the second study, 118 patients with primary diagnosis of advanced nasopharyngeal
carcinomas (stage II-IVb) without distant metastasis were integrated; 88 of them were
examined in a training cohort and 30 in an independent validation cohort. A radiomic
signature could be established by means of a combination of CET1-weighted and T2-weighted
images together with the TNM stage regarding the progression-free survival. This was
superior to a signature of CET1-weighted or T2-weighted images alone and also TNM
classification alone [65]. The significance of radiomics was improved by the combination with known clinical
parameters – or vice versa – in the sense of multimodal modeling.
The team of Lambin et al. may be considered as pioneers in the field of radiomics
in general and in particular of head and neck cancer. Their primary research in head
and neck cancer is based on routinely performed CT scans. In cases of squamous cell
carcinomas of the head and neck, there seem to be similar effects to small cell lung
cancer. In 2014, 440 automatically extracted radiomic features were examined in computed
tomographic scans of 1,019 patients who either had lung or head and neck cancer. They
included phenotypic properties that reflected the tumor image intensity, shape, structure,
and waves in several scalings. The stability of those characteristics was first tested
in 2 small cohorts (n=31 and n=21). A radiomic signature could then be established
in a larger cohort of 422 lung cancer patients (Lung 1/Maastro) that was correlated
with the clinical outcome (Kaplan-Meier diagram) of the patients.
It contains 4 parameters: statistics energy, shape compactness, grey level nonuniformity,
and grey level nonuniformity HLH. For validation, 4 additional cohorts were included.
In 3 of them, independence was already clear because they originated from different
study centers (Lung 2/Radboud n=225, H&N2/VU Amsterdam n=95, Lung 3/MUMC n=89, H&N1/Maastro
n=422). The established signature could be validated in 3 cohorts (Lung 2, H&N1, H&N2).
Astonishingly, it was superior to the predictive significance of TNM staging alone
in Lung 2 and also in H&N2 and comparable to the TNM classification in N&N1. A combination
of the TNM staging with the radiomic signature could further improve the prediction
of the outcome in all groups – independent from the patients’ treatment (radiation
or radiochemotherapy). In particular after publication of the revised TNM classification
that had not been applied in the context of this study, it is additionally interesting
if HPV (human papillomavirus) positive can be differentiated from HPV-negative patients
by means of a radiomic signature, especially because their outcome is different after
radio(chemo)therapy [66]
[67]. However, this was not the case although the outcome could be well predicted by
the radiomic signature in particular in HPV-negative patients. In addition to clinical
data correlation, the radiomic signature of the Lung 3 cohort was correlated with
corresponding genetic data of the same cohort in a gene-set enrichment analysis (GSEA).
Hereby, associations between the expression of different genetic groups and the radiological
structural parameters could be defined. In particular, genetic expression variations
of the cell cycle were depicted. Hence, also the molecular biology on which the tumor
is based can be revealed by imaging up to a certain degree [13] and thus the value of radiogenomics is supported.
The same 440 CT-based radiomic features were applied in another study of the group.
This time they were correlated with further clinical properties. The study was – similar
to the previous one – subdivided into a training and a validation phase. Two cohorts
with lung or head and neck cancer were assigned to a training cohort (Lung 1 n=422,
HN1 n=135) and the signatures were validated in 2 additional independent cohorts (Lung
2 n=225, HN2 n=95). Comparing head and neck with lung cancer, 143 characteristics
were relevant for both tumor entities. In addition, 190 parameters characterized the
outcome only for lung cancer and further 22 radiomic parameters were only relevant
for head and neck cancer. Different clusters could be correlated with survival (lung
and head and neck cancer), histology (lung cancer), and tumor stage (lung and head
and neck cancer), however, the HPV status could not be revealed by a signature [68].
In order to further develop the methods of radiomics, the team established a machine-learning-based
method that could predict the overall survival of head and neck cancer patients based
on a radiomic signature with high stability. The objective was to improve the practical
application of a radiomic signature also for the clinical routine. This is necessary
to introduce radiomics as a non-invasive, cost-effective method in the medium term.
The already known 440 radiomic features were tested by means of 13 methods for characteristics
selection and 11 machine-learning classification methods in a first cohort consisting
of 101 head and neck cancer patients and validated in another independent cohort with
95 head and neck cancer patients. The endpoint was the overall survival. Hereby, a
reliable machine-learning method could be identified [69].
In summary, those 3 publications were able to correlate data of different origins
(clinical, histological, and genetic information) with radiomics parameters and in
this way characterize the tumor based on imaging. The radiomic signature alone was
sometimes superior to single data resources that had been applied for characterization
before. But even when this was not the case, their completion in the sense of multimodal
modeling could improve the significance for the assessment of the carcinoma. The trials
had been performed with relatively large patient cohorts from institutions in partly
different locations so that they were methodically well designed and the reliability
of their significance could already be verified internally.
Recently an article of the “Head and Neck Quantitative Imaging Work Group of the M.D.
Anderson Cancer Cancer/MICCA” was published. In a newly initiated trial, 288 patients
with oropharyngeal cancer and known HPV status were included. They had undergone primary
radiotherapy with curative intention (iMRT) and a standardized pretherapeutic CT scan.
As primary endpoints of the radiomics-based analysis of the CT scans, the HPV status
and the occurrence of local recurrences were defined. In an approach designed as a
competition with scoring, different researchers could test their algorithms regarding
the evaluation of the HPV status and the local recurrences. During the annual meeting
of the MICCAI 2106, the winners were presented [70].
Also for the practical application of radiomics in clinical therapy, first studies
are available for head and neck cancer. Based on CT scans, Ou et al. investigated,
together with the team of Philippe Lambin, 544 imaging characteristics in 120 patients
with advanced head and neck cancer. The patients received radiochemotherapy or “bioradiotherapy”.
Based on pretherapeutic planning CT scans, the overall survival (HR=0.3; p=0.02) and
the progression-free survival (HR=0.3, p=0.01) could be predicted by means of a radiomic
signature encompassing 24 characteristics. A combination with the p16 status as indicator
for the biomarker HPV further improved the significance of the signature. Overall,
this combination was more relevant than the p16 status or the radiomic signature alone
[71]. In another trial, FDG-PET images of 174 patients with advanced stage III-IV oropharyngeal
carcinoma were examined who received definitive radiochemotherapy. Imaging was performed
before and after therapy. As endpoints, the mortality, the local treatment failure,
and distant metastasis were defined. In this investigation, 24 representative radiomic
features were included that reflected the tumor intensity, shape, and structure. Predictive
models for the mortality, the local treatment failure, and the occurrence of distant
metastasis could be established that were cross-validated internally. Unfortunately,
this model did not reach significance for local treatment failure during external
validation. In addition, the models for mortality and distant metastasis could not
be confirmed statistically although, according to the authors, they had an acceptable
predictive performance [72].
Overall, many promising approaches exist for head and neck cancer to usefully establish
and clinically integrate radiomics. Activities of other research teams that enrich
the field with further independent investigations would be desirable. CT scanning
as the standard imaging procedure for head and neck cancer seems to be a reasonable
basis, although further MRI-based examinations still have to be assessed. At the same
time, the available studies of head and neck and other carcinomas show that beside
the data quality and quantity, the success of the studies is fundamentally determined
by a structured approach in the context of radiomics. So, how to approach those data?
3. Practical Implementation of Radiomics
3. Practical Implementation of Radiomics
Meanwhile, a typical radiomics workflow has been defined that is used generally in
nearly all studies.
First, a suitable standardized imaging procedure is identified. Then, the region of
interest is determined, e. g. a tumor, which is then segmented. From the segmented
areas, radiomic features are defined and extracted by means of specific algorithms.
Together with data from other sources, they are included into a database and accordingly
formatted and are then ready for processing. Using suitable statistical methods, biomarkers
and the radiomic signatures can be defined from those databases. This principle seems
to be logical and relatively easy to handle.
However, despite the application of a standardized workflow, each of those steps bears
the risk of errors, difficulties, and limitations that may impair, falsify and complicate
the analysis. The quality of the analysis, its significance and comparability could
potentially be impaired despite the availability of suitable imaging material. Already
minor changes of the standard or the methods may have effects that reduce the reproducibility.
As a consequence, the established radiomic signature would not be stable and applicable.
So always large cohorts with many possibly comparable datasets are preferred as the
basis for establishing a radiomic signature. Validation in an independent cohort –
ideally by an independent group of researchers – according to standardized protocols
is useful in order to minimize internal sources of errors that are sometimes difficult
to identify. Typical sources of errors will be described in the following according
to the single steps of the workflow.
3.1 Imaging
Imaging is the essential basis to practically perform radiomics. It has to be available
as digital file. All imaging included in a study has to be performed based on the
same standard in the same modality and if possible at a comparable stage of the disease.
The suitable, high-quality imaging modality, the appropriate examination protocol,
and the most reasonable ROI have to be identified. Regarding the analysis of solid
carcinomas, the ROI is mainly the tumor, but it may also include the surrounding tissue
or possible metastases. Even specific anatomical structures, disease foci etc. may
be defined. Optimally, a standard imaging is chosen with the standard protocol for
the most common questions.
Different imaging modalities lead to different radiomic features with potentially
different significance and specificity. Depending on the question and the ROI, more
information is expected in terms of a certain modality. However, if this is not the
case, the investigator probably benefits most from the standard imaging of the analyzed
disease where he has most images at his disposition without additional efforts. This
is important because the larger the cohort is, the potentially higher is the statistical
significance and the less errors occur due to outliers. If the modality is selected,
the type of how imaging is performed as well as the parameters of imaging may still
be adjusted. These are very basic factors that nonetheless have to be defined exactly.
The use of different scanners can also have an impact. However, different scanners
may already exist in separate institutions and replacing them is very expensive. Thus,
particularly in the context of multicenter trials the scanner model should be taken
into account. Regarding the acquisition of imaging, different slice thicknesses, programs,
configurations, or details can be selected. Each of those program features has an
influence on the structure of the imaging data. In cases of contrast media application,
the radiomic features vary potentially according to the type and quantity of the contrast
agent, the time of application, and the physiology-related individual distribution
patterns of the patient. Not least, images taken at different times and stages as
well as metabolic conditions of a disease may be acquired. In this way, the definition
of the ROI and its segmentation can be modified. All these parameters influence the
characteristics to be measured. Some of the above-mentioned variable parameters of
imaging are difficult to influence and thus always a potential source of error. So
it is even more important to strictly standardize all those characteristics that allow
standardization. They include in particular technical standards. The type of image
acquisition, the imaging mode, the matrices, the slices, resolutions, reconstruction
as well as the type and adapted quantity of the contrast agent can be widely standardized
allowing comparability between separate study centers. Fortunately, the introduction
of standards for specific questions is increasingly accepted so that this prerequisite
for radiomic analyses is structurally improved. Accurate clinical information about
the type and stage of disease, metabolic diseases, and other clinical features help
in cases of instable variables to identify influences and at least to take them into
account in the context of data processing [73].
3.2 Segmentation
Segmentation defines the borders of the area that is analyzed by radiomics, e. g.
a tumor. So segmentation is an essential step and a basic precondition for performing
radiomics. The region of interest (ROI) and the volume of interest (VOI) are identified.
Slice per slice the ROI is marked in the images so that itself and its relation to
the surrounding structures are finally depicted in a completely 3-dimensional way.
Only structures that are included, are considered later in the analysis. A definition
and segmentation of several different ROIs as well as their later assessment, together
or separated, is generally possible. Radiological features are relevantly influenced
by segmentation. Methodically, segmentation can be generally performed manually, semi-automatically,
or automatically. Up to now, not all 3 methods are available for each application.
The more exact the borders are defined and the better they can be delineated from
the surroundings, the easier is the establishment of a semi-automatic or automatic
segmentation. ROIs that are more difficult to define are often manually segmented.
Since for this purpose, the examiner has to delineate the ROI slice per slice and
its borders must be defined, this process is very time consuming and incompatible
with clinical routine. Manual segmentation should always be performed by a specialist
because its quality is highly dependent on the examiner’s experience. Nonetheless,
there is a high interobserver variability that influences the radiological signature.
Even the same examiner may define segmentations of the same ROI in a different way
at different times, according to the high intraobserver variability. The automatic
segmentation is performed by mathematical algorithms. Due to this gain in objectivity,
the intra- and interobserver variability may be neglected. The segmentations have
a better reproducibility and can be performed more rapidly. Thus, automatic segmentation
is very suitable for large datasets, many datasets, and also multicenter approaches
with many different examiners. However, it is not possible for every ROI. Blurred
object borders, missing clear contrasts in the same narrow localizations or also artefacts
are problematic in the context of automatic analysis limiting its applicability. This
may lead to falsely defined ROIs – even if they are reproducible. The tumor might
for example not be included completely or the software scans alternative areas. It
is worth discussing if an automatic misinterpretation of an automatized segmentation
or the intra- and interobserver variability of manual segmentation are the greater
evil for the application of radiomics. The compromise and at the same time often the
precursor of automatic segmentation is semi-automatic segmentation. Hereby, the ROI
is identified by the examiner and circumscribed for example in a slice of imaging.
The software then performs an automatic segmentation of the defined object and the
examiner post-edits it. Semi-automatic segmentation simultaneously includes sources
of errors and advantages of the manual and automatic approaches. It is more rapid
than manual segmentation but it suffers from the intra- and interobserver variability
of the manual segmentation [73].
3.3 Establishment of radiomic features
The ROI defined by segmentation is analyzed automatically by means of specific algorithms
that compile numeric values by analyzing voxels and pixels. Hundreds of features may
be produced and varied. They describe for example localizations, intensities, shapes,
structures and structural differences, greyscales, color intensities as well as correlations
and relativity values of these features. The selected features should be verified
before application in a study with regard to their stability within the examination
and in different individuals of the study.
The ROI properties may be depicted in different ways and further processed. So ROI
intensities may be visualized via a histogram that is based on fractionated volume
data on voxel level. Data for example on the form of the ROI (volume, shape, surface
markers, density etc.) provide additional accompanying statistical values. The analysis
of additional, secondary qualities, clusters, correlations – even beyond different
image settings – may provide enormous amounts of data. One challenge is the exclusion
of redundancies. Of course those large data quantities are difficult to handle regarding
their processing. With the help of statistical methods and machine-learning, the parameters
have to be reduced to the informative and validated features that are relevant for
the objective of the analysis and trial. Only in this way, the features gain their
specific significance.
3.4 Establishment of databases
One particularity of radiomics is to reasonably analyze the radiological characteristics
in the clinical, genetic, and/or histopathological context. For this purpose, according
databases have to be compiled. Large data storage capacities that can be well accessed
for analysis have to be available. The definition of the feature value should correspond
to a selection of the possible variables that can be exactly delineated. For each
clinical, genetic, or histological property, the type and source should be the same.
Linking different databases, e. g. clinical, genetic, and radiological ones, may also
be useful. However, the regulations of data protection should be strictly observed.
Of course, all these data have to be digitally available for statistical evaluation.
3.5 Analysis of databases
With the compilation of the databases, the actual evaluation starts. The objective
is the establishment of a radiomic signature that correlates with a specific requirement
or question. An alternative objective may be a multimodal modeling where the radiomic
signature is evaluated together with other data which leads to an additional value
regarding precision and/or information. A radiomic signature is extracted from all
measured radiological values and may contain only one or several features at the same
time. Those features may be trivial in the sense of already macroscopically known
phenomena but also consist of features that are more abstractly gained from voxels
and pixels. There are radiomic signatures that contain a 3-digit number of single
components. Taking into account that already hundreds of values of radiological characteristics
are available per patient and further features are added or that large data volumes
from several sources are processed simultaneously (e. g. combined with genome analysis),
it is clear that experienced statisticians and a suitable software are essential.
The software should be able to assess large data quantities in a reasonable timeframe
and generate solid, reproducible, broadly applicable biomarker signatures. Of course,
thorough approaches during the previous steps significantly improve the data quality.
At that time, sources of errors can only scarcely be limited by analysis and statistics.
A high number of well-defined datasets, possibly obtained from several centers by
different examiners, limits errors due to outliers, interobserver variability, local
particularities, and measurement uncertainty [73]. It is rather logistically difficult and cost-intensive to achieve this objective.
Ideally, a primary radiological signature is created based on a retrospective cohort,
validated in another independent cohort, and tested prospectively in the clinical
setting [13].
4. Radiomics: Study Objectives
4. Radiomics: Study Objectives
In oncology, radiomics are currently used in particular for characterization of solid
tumors on different levels (histological, genetic, clinics-associated), for prediction
of the outcome, and for prediction of the therapy response in the context of primarily
conservative therapeutic measures. A transition into other reasonable correlations,
however, seems to be possible.
Current tumor characterizations encompass clinical data, macroscopically assessable
imaging data, rarely functional data (e. g. mobility of the vocal folds in cases of
laryngeal cancer), and genetic, proteomic, and (immuno)histological information from
tumor areas that are pretherapeutically gained from biopsies. The whole tumor can
only be examined after surgical extirpation. Biopsies reflect a representative tissue
image that may be characterized histopathologically and/or molecular-biologically.
Unfortunately, however, many carcinomas are not homogenous. Various cell populations
and clones with different histological and in particular molecular properties are
found in different areas. Before biopsy, they cannot necessarily be differentiated.
The biopsy area is determined by clinical factors, as its anatomical position and
accessibility as well as the biopsy method and the capacities and experiences of the
examiner who should succeed in obtaining more or less representative tissue. Studies,
where biopsies were taken from different sites of the tumor, already revealed this
phenomenon.
Radiomics have the advantage that they assess the whole tumor by means of morphological
imaging with regard to its size, shape, surface and internal structure as well as
the anatomical context. If the detection of radiomic signatures is successful for
histological, genetic, and proteomic conditions of the tissue, they may potentially
be assessed by radiomics in the entirety of the tumor. Furthermore, radiomics may
be extended to the surrounding tissue and metastases or metabolic conditions. Alternatively,
radiomics might be used to better distinguish subareas of the tumors that cannot be
seen with the naked eye and thus contribute to improve the quality of biopsies and
to indicate where biopsies should ideally be taken in order to histologically and
molecularly characterize the tumor. However, it would be desirable to completely avoid
invasive biopsies. So it would be beneficial to improve the radiomic specificity to
that extent that their characterization is equivalent to the quality of biopsy or
even provides better results. If and for which applications this may be possible,
will have to be proven in further studies. Up to now, radiomics should be used in
the context of clinical data and data of additional sources.
Some studies have already shown that an amalgamation of radiomic features with biomarkers
and data from other sources may achieve a better and more accurate subtyping of diseases
and a better quality of outcome predictions. Until now, information from imaging are
only indirectly included in tumor typing. They contribute for example to the TNM classification.
In the context of the sometimes very rough radiological characteristics that were
commonly considered such as tumor size, invasion, extracapsular spread etc. informative
metadata from the background remain disregarded. Many of them get “lost” for the human
examiner during processing of the images. So they may currently be considered as “dead”
data source. In summary with the data of other sources, radiomics may refine the typing
by other biomarkers and reduce grey areas. Hereby, radiomics could become an integral
part of multimodal prediction or typing models that are finally the basis of a broad
spectrum of applications in the context of personalized medicine, which could be enriched
and improved by radiomics. Because of the enormous data volume and in the context
of a possible standardized objective description, the development and application
of suitable software is essential.
Reasonable outcome correlations could contribute to the assessment of the tumor aggressiveness.
As part of a multimodal overall assessment as well as for specific applications as
potential biopsy substitute, radiomics might improve the prediction and monitoring
of the outcome as well as treatment options.
Regardless of the predictive character, radiomics-based examinations also contribute
to increasingly automatizing cancer screening in general and to improving the associated
standardization. They foster cost- and time-effective diagnostics. The role of the
diagnosing examiner could be either supported or be pushed into the background. However,
it must not be forgotten that radiomic algorithms assess information that the human
eye is not able to realize during visualization of the imaging. In this way, radiomics
even possess the potential to achieve better diagnoses than humans.
And the major advantage is that the required data are already present in our standard
imaging procedures in high quantities. They just wait for being explored. If relevant
radiomic features are defined and validated, additional examinations will no longer
be necessary.
5. Factors for Clinical Integration of Radiomics: A Future Vision
5. Factors for Clinical Integration of Radiomics: A Future Vision
So what will be a future vision? Until now, radiomics in the sense of the current
definition have not been introduced into clinical routine. This is mainly due to their
newness. But they are increasingly recognized as alternative source of biomarkers
and promoted on the scientific level. With further intensification of big data and
IT-based model approaches, imaging might complete or replace information obtained
by biopsies (histological, genetic/molecular). Uni- or multimodal modeling could precisely
predict the outcome and therapy response so that it does not only support medical
decision but even replaces it in extreme cases. Those approaches are only possible
with very strict and multicentric standardization of diagnostics and therapy. Treatment
individualization would be moved forward due to objective data. However, this also
bears the risk of losing the individualization because psychological, social, or generally
spoken “human” characteristics would not be assessed. The physician – and thus also
the patient – would be more and more subordinate to standardization and the power
of the data situation, with all its advantages and disadvantages. A further development
of radiomics and modeling would support telemedicine and self-diagnosis and thus the
centralization of medical service in specialized centers.
But what will have to happen to make radiomics practicable for clinicians? What are
the wishes of the applying physician?
It is necessary that the user may rely on a constantly high specificity and sensitivity
of the analysis because therapeutic and diagnostic decisions depend on it. Ideally,
radiomics should have a broad application spectrum. Based on cost-benefit calculations
and the patient comfort, the use of standard imaging without additional efforts would
be desirable. It should be possible to integrate the respective software into the
local IT systems and thus allow smooth processing when linking clinical data with
genetic and (processed) imaging data, if necessary. The segmentation should be automatized
and integrated into the processing software. It should be user-friendly, i. e. intuitive
and clear. The results of radiomic analysis are better accepted when they can be adequately
visualized. Practically, radiomics may not only reduce or avoid biopsies but additionally
allow for a more holistic assessment of the tumor. A link to the completing software
programs, multimodally oriented outcome prognoses, or therapy concepts, would be further
possible options.
6. Conclusion
Radiomics enlarge the field of biomarkers in an innovative way and the basic data
of imaging that hereby gain in importance are included in the wide spectrum of “omics”
and biomarkers. They could substantially contribute to personalized medicine. A major
advantage is that the data generally already exist and “only” have to be evaluated.
Another advantage is that they may be retrieved without biopsies and their potentially
complex and expensive assessment (e. g. genomics). Nonetheless, hereby an overall
assessment of the tumor is generated and not only an excerpt due to a biopsy. Radiomic
signatures could possibly serve alone as biomarkers and replace other clinical, histopathological,
and genetic markers. In this way, the patient comfort might be improved and financial
means may be saved. An additional benefit may also be generated by multimodal modeling,
correlation with data from other resources and thus extend and improve their significance.
For both purposes, however, it is necessary to process large data volumes, which requires
a high expertise and bears the important risk of potential errors on all levels of
establishment and validation. For clinical integration, not only a high measure of
standardization is necessary but also the implementation of suitable segmenting and
analyzing software that make the definition of radiomic signature realizable in clinical
routine.
6.1 Big data instead of biopsy?
In the future, it might be that radiomics replace biopsies for specific questions.
However, in the very near future it seems to be more probable that radiomics complete
the findings of biopsies and that data models enriched by radiomics improve precision
medicine.
