Parameter estimation of perfusion models in dynamic contrast-enhanced imaging: a unified framework for model comparison
Graphical abstract
Introduction
Providing earlier assessment of drugs efficiency is a major challenge for the improvement of patient care in oncology. Follow-up of patients presenting tumours is traditionally performed through various morphological measurements (number of tumours, size of tumours,...) (Shanbhogue et al., 2010). However, such measurements are known to reflect only partially the disease progression. Indeed, morphological assessment may be insensitive to or provide markedly delayed indications of the tumour response to treatment even when the therapeutic effect is substantial. It has been shown that the physiological response to a given therapy often precedes the evolution of its morphological descriptors (Li and Padhani, 2012). In this context, dynamic contrast-enhanced (DCE) imaging is a promising tool for the assessment of tissue differentiation based on its intrinsic nature (normal, tumour, necrotic...). The technique, based on the temporal analysis of the contrast intake curve extracted from the images after bolus injection of a contrast agent, aims at characterising tissue microcirculation and microvascularisation.
Functional imaging of the microcirculation, using either magnetic resonance imaging (MRI) (Sourbron and Buckley, 2012), computed tomography (CT) (Miles, Lee, Goh, Klotz, Cuenod, Bisdas, Groves, Hayball, Alonzi, Brunner, 2012, Ingrisch, Sourbron, 2013), ultrasound (Lassau et al., 2010), positron-emission tomography (Keiding, 2012) or single-photon emission CT (Zhang et al., 2012), is based on dynamic contrast imaging, although the diffusion kinetics depends on the type of contrast agent. In this paper, we focus on DCE-MRI and DCE-CT. In such cases, contrast agents are small inert molecules that propagate in extracellular spaces. The kinetics of contrast agent diffusion are similar for both modalities.
From contrast intake curves, simple parameters, such as time to peak or mean transit time (Našel et al., 2000), may be extracted to describe the change of contrast agent concentration and compute parametric images. However, more informative approaches exist, that aim to provide physiologically-based parameters. By associating the temporal variation of contrast intake to physiological parameters, the analysis of contrast intake curves provides a way to characterize the pharmacokinetics of the contrast agent. These parameters can be derived using the tracer kinetic theory (Brix et al., 2010). In this theory, two different concepts have been highlighted. The first one, the indicator dilution theory, is based on a convolution approach and does not rely on any assumption about the diffusion process (Meier and Zierler, 1954). The second approach, which will be considered in this paper, is based on pharmacokinetic compartment models, usually formulated as systems of coupled differential equations. They rely on the assumption that tissues can be represented as a set of interacting subcompartments within which an administered tracer can circulate with different dynamics that depend on each compartment properties. Compartment models are likely to provide a better biological insight, relying on several physiological parameters such as the local vascular permeability, blood flow, intravascular and extracellular volumes (Ingrisch and Sourbron, 2013). These parameters can be extracted from contrast intake curves in a given tissue region through model parameter estimation. However, several models have been developed in the literature and no consensus exists on their clinical use. Each type of model has long been confined to specific applicative domains and only recently efforts have been undertaken to review the full range of existing models (Sourbron, Buckley, 2012, Ingrisch, Sourbron, 2013, Brix, Griebel, Kiessling, Wenz, 2010). For instance, Sourbron and Buckley (2012) and Ingrisch and Sourbron (2013) present a classification that provides clearer insight into the links between the different models. Ingrisch and Sourbron (2013) discuss the difficult problem of model selection: they argue that, while the classical goodness-of-fit criteria such as the χ2 value are the most widely used, they are inappropriate to compare models with different numbers of free parameters and should be replaced in that case by penalized criteria such as the Akaike information criterion (AIC). Some mathematical aspects of model analysis, and their systematic applications to a whole group of models, are nevertheless still lacking. Indeed, to obtain reliable and relevant estimates of parameters, one must ensure that the chosen model meets a certain number of requirements: plausibility of the underlying hypotheses and parameters biological meaning, parameter sensitivity and identifiability, solvability of the optimization problem associated to the estimation process, robustness to noise. With a few exceptions (Orton et al., 2007), these conditions are hardly examined; most studies focus on direct applications of the models without questioning their validity.
Therefore, complementary to other review papers that present and classify the models, the objectives of this paper are (i) to define the biological context and general form of perfusion models, (ii) to propose a generic framework for parametric estimation problem of perfusion models in the context of DCE-MRI or DCE-CT imaging through the definition of a unified mathematical framework providing tools to perform robust parameter estimation, sensitivity and identifiability analysis, (iii) to apply the framework to the most widely used compartment models, and finally (iv) to illustrate the results on abdominal DCE-CT images.
Section snippets
Data
DCE imaging consists in the acquisition of temporal sequences of 3D images. Several acquisition protocols can be selected, depending on the modality (CT or MRI) (Kambadakone, Sahani, 2009, Sourbron, 2010) and on the values of the temporal resolution (time delay between consecutive acquisitions) and the duration of acquisition. In this work, the parameters of reconstruction (the spatial resolution and the smoothing filter) were set following the recommendations of Romain et al. (2012). From
Description of extracellular models
The most general model is the Distributed Parameter (DP) model. First introduced by Sangren and Sheppard (1953), it makes no additional assumption on the two-compartment structure than those presented in section 2.2 and uses the four parameters: . The impulse residue function R(t) for the DP model is composed of a vascular transit phase Rvasc(t) and a parenchyma phase Rpar(t): with denoting the vascular transit time, i.e. the mean time for blood
Loss function and constraints
In the optimization problem described in Eq. (7), the goal is to minimize the cost function i.e. in the least squares case, with respect to the constraints. Some of the constraints are common to all models (e.g. positivity of the parameters), and a maximum of three additional constraints may be added, depending on the involved parameters: and .
Given the non-convexity of objective functions associated to the five selected models,
Discussion
In this paper, we proposed a methodology for model assessment in the context of parametric estimation from DCE-CT or DCE-MRI imaging. This methodology gathers a set of mathematical and statistical tools that we recommend to systematically apply before using a model in a particular clinical context. Few other studies have focused on model comparison. Ingrisch and Sourbron (2013) discussed the problem of model selection and recommended the use of the AIC to choose, for each clinical application,
References (56)
- et al.
On impulse response functions computed from dynamic contrast-enhanced image data by algebraic deconvolution and compartmental modeling
Phys. Med.
(2012) - et al.
Perfusion estimated with rapid dynamic contrast-enhanced magnetic resonance imaging correlates inversely with vascular endothelial growth factor expression and pimonidazole staining in head-and-neck cancer: a pilot study
Int. J. Radiat. Oncol. Biol. Phys.
(2011) - et al.
Body perfusion CT: technique, clinical applications, and advances.
Radiol Clin North Am
(2009) CT hepatic perfusion measurement: comparison of three analytic methods.
Eur J Radiol
(2012)- et al.
Inverse modelling, sensitivity and monte carlo analysis in r using package fme
J. Stat. Software
(2010) - et al.
Tracer kinetic modelling in MRI: estimating perfusion and capillary permeability.
Phys Med Biol
(2012) A new look at the statistical model identification
Autom. Control, IEEE Trans.
(1974)- et al.
Systems biology: parameter estimation for biochemical models
Febs J.
(2009) - et al.
Comprehensive model for simultaneous MRI determination of perfusion and permeability using a blood-pool agent in rats rhabdomyosarcoma
Eur. radiol.
(2005) - et al.
Dynamic contrast-enhanced CT of head and neck tumors: perfusion measurements using a distributed-parameter tracer kinetic model. initial results and comparison with deconvolution-based analysis
Phys. med. biol.
(2007)
Regional blood flow, capillary permeability, and compartmental volumes: measurement with dynamic ct?initial experience
Radiol.
Tracer kinetic modelling of tumour angiogenesis based on dynamic contrast-enhanced ct and mri measurements
Eur. j. nuclear med. mol. imaging
Pharmacokinetic analysis of tissue microcirculation using nested models: multimodel inference and parameter identifiability
Med. phys.
Practical identifiability analysis of large environmental simulation models
Water Resour. Res.
Model selection and multimodel inference: a practical information-theoretic approach
The microvascular characteristics of cervical cancer: limitations of the modified-tofts tracer kinetic model for the analysis of DCE-MRI data
Proc Intl Soc Mag Reson Med
Interior methods for nonlinear optimization
SIAM rev.
A novel estimation method for physiological parameters in dynamic contrast-enhanced MRI: application of a distributed parameter model using fourier-domain calculations
Med. Imaging, IEEE Trans.
Quantitative tumor perfusion assessment with multidetector CT: Are measurements from two commercial software packages interchangeable? 1
Radiol.
Tracer-kinetic modeling of dynamic contrast-enhanced MRI and CT: a primer.
J Pharmacokinet Pharmacodyn
A model for capillary exchange
Am. J. Physiol-Legacy Content
Bringing physiology into PET of the liver.
J Nucl Med
Fundamentals of tracer kinetics for dynamic contrast-enhanced MRI
Journal of Magnetic Resonance Imaging
Dynamic contrast-enhanced CT imaging of hepatocellular carcinoma in cirrhosis: feasibility of a prolonged dual-phase imaging protocol with tracer kinetics modeling
Eur. radiol.
Hepatic metastases: In vivo assessment of perfusion parameters at dynamic contrast-enhanced MR imaging with dual-input two-compartment tracer kinetics mode
Radiol.
Tracer-kinetic models for measuring cerebral blood flow using externally detected radiotracers
J. Cereb. Blood Flow Metab.
Dynamic contrast-enhanced ultrasonography (DCE-US): a new tool for the early evaluation of antiangiogenic treatment
Targeted oncol.
Cited by (15)
Physics-informed neural networks for myocardial perfusion MRI quantification
2022, Medical Image AnalysisCitation Excerpt :As a result, the model-based concentration curves may match the noisy observed data, but the inferred parameters can be far from the actual values. It has further been shown that model parameters are correlated (Romain et al., 2017), which causes several distinct parameter combinations to produce concentration curves which may all well fit the observed data, therefore making the identification of the correct set of parameters difficult (Buckley, 2002; Ahearn et al., 2005). This has lead to more complex fitting algorithms being proposed but these have not yet seen widespread adoption (Kelm et al., 2009; Dikaios et al., 2017; Scannell et al., 2020a; Dikaios, 2020).
Reproducibility of Computed Tomography perfusion parameters in hepatic multicentre study in patients with colorectal cancer
2021, Biomedical Signal Processing and ControlCitation Excerpt :Many studies report variations of up to 30% between perfusion values, depending on the computing methods chosen [20]. On the other hand, very few methodological studies deal with how to improve CTp reproducibility and even less [21] focus on the modelling aspects rather than on the computational ones. In this regard, Deconvolution (DV) and Maximum Slope (MS) are well-established and widely used perfusion methods, independent from each other.
Hierarchical Bayesian myocardial perfusion quantification
2020, Medical Image AnalysisCitation Excerpt :As a result, even though the model-based concentration curves may well match the observed data, the reported parameters may be far from the true values. It has further been shown that the model parameters are correlated (Romain et al., 2017) and thus there are multiple distinct combinations of parameters that give outputs that are indistinguishable at the observed noise level. Also, as is typical with non-linear optimisations, the parameter estimates are highly sensitive to the initial conditions of the optimisation and the specific noise present in the data.
A Novel Statistical Optimization Algorithm for Estimating Perfusion Curves in Susceptibility Contrast-Enhanced MRI
2021, Frontiers in Neuroscience