Radiomics workflow: steps and challenges
The Radiomics workflow includes multiple steps: data selection and image acquisition; image segmentation; features extraction (pre-processing steps); exploratory analysis/feature selection,
model building and validation [3-4].
Data Selection and Image acquisition
A Radiomics analysis begins with the choice of an imaging type,
the region/volume of interest (ROI/VOI) and a prediction target.
Medical images acquired for standard clinical diagnostics,
treatment planning and follow-up can be used,
with the possibility to obtain a big amount of data [4].
A principal drawback is that there are wide variations in images acquisition and reconstruction protocols,
both among exams in the same hospital or in different centres [4].
When images are analysed quantitatively any changes in technical parameters can lead to changes in the extracted data which are not correlated with biological factors,
introducing confounding variability or the so-called “noise”.
The use of different scanners,
slice thickness,
reconstruction algorithms or parameters as well as the standardized-uptake-value (SUV) discretization can influence the repeatability and robustness of the Radiomics features [4].
In addition should be taken into account the influence of patient’ related factors,
especially the respiratory motion.
Phantom studies can identify features that rely on the vendor,
when data are derived from multiple scanners [3].
Test-retest data (e.g.
datasets of images acquired within a small period of time or at different stages of the breathing cycle) can be applied to measure Radiomics feature stability [3].
Finally specific Radiomics correction and calibration algorithms could potentially solve the variability of the features obtained from retrospective studies [4].
For future prospective analyses,
however,
the use of standardized imaging protocols would be desirable [3].
Image segmentation and rendering
Imaging features are derived from a defined ROI/VOI of interest,
therefore are highly dependent from segmentation [4]: the variability in segmentation can introduce bias in the evaluation of Radiomics features.
Manual delineation is commonly used in clinical routine,
for example in the delineation of radiotherapy treatment planning [4].
However it is time consuming and is susceptible to inter-observer variability [4].
Automatic or semi-automatic segmentation methods by minimizing manual input will outperform manual segmentation in terms of repeatability [4].
Nevertheless their performance is still depending in the methods or software used [4].
Multiple segmentations (e.g.
evaluation by multiple clinicians,
perturb segmentations with noise) can limit the above-described bias [4].
Features extraction and qualification
Multiple classes of quantitative imaging features can be used,
both morphological and textural.
The term “image texture” refers to the perceived or measured spatial variation in the grey levels [2]: texture analysis offers an objective quantitative assessment of tumour heterogeneity by analysing the distribution and relationship of pixel/voxel grey levels.
It is performed by different methods: statistical (most used),
model-based or transform-based.
Morphological features. They are derived from 3D geometrical rendering of the segmented volume [4].
They describe:
-Size: volume and maximal diameter
-Shape: sphericity,
compactness and surface/volume ratio.
Finally,
features relating to tumour location (e.g.
involved lung lobe) have been implemented.
Textural features.
- Statistical.
They are obtained from the distribution of the grey values in the frequency histogram and their spatial relationship [2,4].
The derived parameters are classified according to the increasing complexity of the spatial distribution described: first order features are independent form the spatial dimension,
by the contrary higher order ones describe the spatial arrangements of the intensities [2,4] (Figure 2).
The statistical features are described and represented in Table 1 and Figure 3,
respectively.
- Model-based.
They are based on mathematical models.
One of the main used is Fractal analysis,
which investigates repetitive patterns at different level of complexity in the image.
The fractal dimension (FD) describes the complexity of the object [2].
- Transform-based.
They are obtained from the transformation of the spatial information into frequency (Fourier) or into scale and frequency (wavelet) information [2].
A pre-processing step can be done before features extraction,
by applying [4]:
- Filters that reduce noise (e.g.
Gaussian filter) or imaging enhancing techniques (Laplacian filter,
histogram equalization,
de-blurring and resampling);
- Discretization of the grey values in defined ranges or “bins”,
influencing both sensitivity and reproducibility of the features;
- Resampling voxel size to an anisotropic voxel (matrix-based textural features are highly sensitive to the voxel size anisotropy).
Radiomics features can be calculated from one slice of the ROI/VOI,
considering the one with largest cross-sectional area,
or in a volumetric way by taking into account all the voxel within the ROI/VOI [2].
The use of volumetric system is recommended in order to include the whole tumour [2].
Variation exists in features nomenclature,
mathematical definition and methodology of extraction [1].
Moreover multiple software packages are actually available to perform texture and Radiomics analysis [2-3]. Beside to the elucidation of the methodological aspects and software used,
research projects are on-going to offer standardized settings for image processing [3-4].
Data analysis and model creation
Exploratory analysis/feature selection
The number of the features that could be extracted from images is usually larger than the number of the patients involved [3-4]: it leads to a possible over-fitting of the features,
resulting in “spurious” relationships [3-4].
Dimensionality reduction strategies therefore need to be applied.
Features that are prone to variability should be identified and eliminated (e.g.
phantom studies,
multiple images test-retest,
multiple segmentation) [3-4].
First order features are more stable,
especially kurtosis and entropy.
Most of the imaging features will show some correlation,
as they derive from the same matrix.
Filter-based selection techniques select informative ones as well as identify redundancy,
highlight outliers.
Clusters of highly correlated features are identified and reduced in one representative (“archetypal”),
usually the one less susceptible to source of variability or the most informative [3-4].
Modelling and validation
Data from multiple disciplines (e.g.
clinical records,
genetic profiling) besides images are finally included,
as they potentially influence not only the outcome variable but also the extracted Radiomics features [3-4].
The aim of most Radiomics analyses is to obtain a prognostic or predictive model,
efficient and accurate [3-4].
Modelling is usually performed by Machine-learning (ML) methods [3-4].
ML classifiers are distinguished in supervised (e.g.
linear models,
random forests,
support vector machines and neural networks) or unsupervised,
if they separate the data with respect to an outcome variable or not [4].
Each technique has his own limits (e.g.
the independence assumption for features in logistic regression,
the need for feature discretization in Bayesian networks,
the network configuration dependency in deep learning) therefore the methodology chosen can affect the final performance of the model [3].
Multiple-modelling methodology implementations will be desirable [3-4].
Validation techniques are applied to assess model performance [3-4].
The model is trained on a dataset and then validated in the same population by using mathematical methods (internal validation) or in a different one (external validation): the last system is preferred because more independent,
however requires huge amount of data [3-4].
Cross-validation has frequently been used as internal validation [3].
Valid models should exhibit statistical consistency between the training and validation sets [3-4].
Radiomics application
In the context of lung cancer the main imaging techniques to consider are computed-tomography (CT) and positron-emission-tomography (PET-CT).
Multiple are the possible field of interest of quantitative image analysis: a) Histopathologic and biologic features: correlating with phenotype and genetic; b) Lesion Characterization; c) Primary tumour assessment: predicting Tumour Aggressiveness and Clinical Outcome.
Histopathologic and biologic features: correlating with phenotype and genetic (Table 2)
Tumour’s microenvironment is composed by different tissue types (cancer cells,
stroma and extracellular matrix) and it is the result of continuous changes driven by interaction local factor,
as hypoxia.
The interaction activates “molecular pathways” in cells (like p53,
E-cadherin,
hypoxia inducible factor–1 (HIF-1) alpha,
glucose transporter 1 GLUT-1,
increased glycolysis and CD34 expression),
triggering processes such as cell division and inflammation as well as selection of cancer clones [2,5-6].
All these biological processes play key roles in lung cancer aetiology,
in the emergence of drug resistance and tumour metastasis [2,5-6].
As it is supposed that imaging structure may be determined by the components of the tumour microenvironment at multiple scales,
tumour phenotypic characteristics may be determined by Radiomics analysis.
Specific genetic patterns and mutation are associated with different prognosis and could be responsive to personalized therapy (e.g.
K-ras oncogene,
EGFR,
ALK/ROSE/RET mutation) [8-10].
The aim of “Radiogenomics” is to correlate the imaging analysis with the genetic one,
replacing the invasive histological biopsy with a “virtual one”.
Lesion Characterization (Table 3)
An accurate non-invasive lesions classification is determinant for the adequate management,
regarding both lung nodules and lymph nodes [2].
Promising results showed that texture differences exist among benign and malignant lesions,
the last ones generally with greater heterogeneity and irregular shape [2].
Regarding lung nodules a potential role of Radiomics analysis has been demonstrated for lesions characterization in the routine clinical setting as well as in a screening scenario; a limit is that most studies concerning nodule classification have been performed using small samples and have not compared texture analysis with existing imaging standard [2,11-12] (Figure 4-5).
Both CT and PET-CT have limited accuracy in characterization of lymph nodes,
especially when they are of small dimensions.
Studies showed that textural analysis might be accurate in diagnosing lymph node metastasis at least as morphologic (short-axis diameter) and metabolic measures (SUV max); interestingly the developed models correctly classified sub-centimetre lymph nodes [2,13-14].
A drawback for texture analysis of lymph nodes is that they are relatively small structures,
therefore limited information can be contained in few voxels [2].
Primary tumour assessment: predicting Tumour Aggressiveness and Clinical Outcome (Table 4)
Different authors suggested that Radiomics parameters could be prognosticator independent tumour stage and clinical factors.
Generally tumour heterogeneity as defined by quantitative image analysis has been related to the Patients prognosis,
moreover the definition of “image phenotype” can stratify patients with different survival [2,
15-16] (Figure 6-7).
Tumour heterogeneity before therapy and its change during the course of treatment may be related with therapy response,
therefore quantitative imaging analysis may be useful in treatment planning and prognostication [2,
17-20].