Practical Aspects of Deconvolution
Fernando Calamante
Matus Straka
Lisa Willats
Introduction
As described in Chapter 24, the tracer kinetic model used to quantify perfusion is given by a mathematical convolution expression. This model applies to both dynamic susceptibility contrast magnetic resonance imaging (DSC-MRI) and perfusion computed tomography (CT; Nb: In fact, the convolution model is also applicable to arterial spin labeling perfusion MRI—see Chapter 31); the main difference relates to the way the concentration of the contrast agent is measured. Therefore, most of the statements described in this chapter regarding DSC-MRI are also relevant to CT perfusion quantification.
Quantification of perfusion requires measurement of the arterial input function (AIF), which is the function describing the input of the contrast agent to the tissue (see Chapter 26). Perfusion is then obtained by inverting the convolution expression (a mathematical process known as “deconvolution”), which removes, from the tissue concentration time course, the temporal spread contribution associated with the AIF. This chapter focuses on (a) the practical issues regarding this deconvolution step, (b) the factors that need to be taken into account, and (c) how it can be used to obtain absolute measurements of perfusion and other hemodynamic parameters.
What is Deconvolution?
The measured tissue tracer concentration depends not only on cerebral blood flow (CBF) but also on the way the contrast agent enters the tissue. In fact, two tissue areas (or two patients) could have exactly the same tissue concentration curves but very different perfusion status: In one case, the bolus could arrive very fast and with a narrow shape but then have a slow clearance owing to severe hypoperfusion; in the other case, the bolus could arrive very slowly and with a spread shape but could be cleared very quickly owing to normal perfusion or hyperperfusion relative to the previous patient. The two situations are obviously very different, but one cannot distinguish them based on the tissue concentration changes over time alone. To differentiate them, the contribution of each patient’s AIF must be taken into account; this is the role of the deconvolution step. Because of its mathematical nature, deconvolution is often regarded as an obscure step by clinicians and nontechnical users. Although many technical aspects can indeed be very complex, the general concepts underlying the deconvolution step can be easily understood by a simple analogy: The audience attending a lecture. (See Geek Box 1 for a more technical description of deconvolution.)
Geek Box: What is Deconvolution? A Technical View
The purpose of deconvolution is to deliver quantitative parameters, which are derived from the response of the system to a standardized input. Such standardization is necessary because the system’s output is always a convolution of its internal dynamic properties with the input. By relating to a standardized input or removing the effect a variable input can have on the output, processing results become comparable between subjects or between subsequent measurements in one subject. In perfusion measurements, the AIF cannot be assumed to be always the same; deconvolution facilitates computation of the system’s output as a response to some standardized (but hypothetical) input. Under the assumption that the investigated system conforms to linear time-independent systems theory, its dynamic properties describe how slowly/quickly varying signals (i.e., low and high frequencies) pass through that system, and whether there is any attenuation and/or delay observed for a given frequency. Summaries of these attenuations and delays at all frequencies (spectral properties of the system) allow a quantitative comparison of dynamic systems. Signals with spectra identical to the spectral properties of the system would be observed on its output after a perturbation with a narrow impulse (i.e., Dirac delta). Therefore, the time-domain representation of the system’s spectral properties is called the impulse response function (IRF). Deconvolution computes the IRF from the spectra of the system’s output and input.
In perfusion measurements, the deconvolution computes the response of the tissue vasculature to a hypothetical very narrow bolus of the tracer. The response (i.e., the IRF) is also called the (scaled) residue function, as the temporally resolved values of the IRF represent the amount of tracer still residing in the hemodynamic system over time, after such hypothetical narrow-bolus injection. Comparison of the residue function properties
(amplitude, width, area under the curve, shape, etc.) then allows direct quantitative comparison of hemodynamics in different tissues (see the second Geek Box), regardless of the specific injection protocol and/or patient cardiac output.
(amplitude, width, area under the curve, shape, etc.) then allows direct quantitative comparison of hemodynamics in different tissues (see the second Geek Box), regardless of the specific injection protocol and/or patient cardiac output.
Mathematically, the convolution is denoted by (see Chapter 24):
where cA(t) is the AIF, CBF·r(t) is the IRF or the (scaled) residue function, ⊗ is the convolution operator, and cT(t) is the concentration time curve (CTC) in the tissue. Here, “scaled” is referring to the multiplication of the residue function, r, with CBF. To obtain the unknown scaled residue function, an inverse operation (deconvolution) must be executed:
where ⊗−1 indicates the deconvolution operator. Solving for the inverse of this integral equation analytically is in general not practical, and numerical approximations are typically used instead. For such task, discretized time- or frequency-domain approximations of the convolution and deconvolution are used.
Time-Domain Formulation
Since CTCs are sampled at TR time intervals (t0, t1, t2, …, tn−1), Eq. [1] can be written in matrix form.1 In particular, assuming that CBF·r(t) and cA(t) are constant over the sampling intervals, Eq. [1] can be approximated by (see the third Geek Box for other approximations):
In shorthand vector-matrix notation this may be written as cT = AIF·(CBF·r), where matrix AIF has elements AIFij determined by the repetition time (TR) and sampled cA(t), and vector CBF·r contains the unknown discretized IRF, CBF·r(t). The matrix-vector multiplication corresponds to the standard notation; for example: cT(t0) = AIF11·(CBF·r1) + AIF12·(CBF·r2) + … + AIF1N·(CBF·rN). To solve for the unknown CBF·r, deconvolution inverts this expression: CBF·r = AIF−1·cT, which requires inversion of the matrix AIF. For real-world signals, however, a simple matrix inversion might not lead to physically plausible solutions for r. Pitfalls of this inversion are discussed later in this chapter.
Frequency-Domain Formulation
Fourier transform theory teaches that convolution in the time domain is equivalent to multiplication in the frequency domain:
where CT(f) = FT[cT(t)], CA(f) = FT[cA(t)] and R(f) = FT[r(t)], and FT is the discrete Fourier transformation. Deconvolution is then the inversion of this equation:
and the IRF can be then obtained as CBF·r(t) = IFT [CBF·R(f)], where IFT is the inverse discrete Fourier transformation. It should be noted that the 1/CA(f) operation becomes ill-conditioned for components of CA(f) → 0, that is, for AIF frequency components with very low power. (Note: In practical terms, an ill-conditioned problem is such that a small amount of noise in the measurement can lead to major errors in the computed solution of the problem.) How such cases are treated is discussed later.
Audience analogy: For an ideal 60-minute lecture, the audience number would be a step function (Fig. 25.1A): Everyone present until the 60-minute mark, and no audience afterward. In practice, however, the audience is more likely described by Fig. 25.1B: Some may leave the room before the end, and some others may stay on to talk with the speaker. Importantly, the shape of this function is an intrinsic property of the lecture, a measure of its “quality”: The worse the lecture, the less people will remain in the audience, with a step function (Fig. 25.1A) representing a very good lecture (and a delta function at time t = 0 the extreme of a very bad lecture!). It is important to note, however, that Fig. 1A and B assume an “instantaneous” audience into the room; that is, everyone arrived at t = 0. In practice, this is never the case, and some arrive late, after the lecture has started. To investigate this effect, consider a lecture theater with one person at every door counting the number of people entering the room. After the doors are opened at t = 0, the total number of people entering the room would look like the function in Figure 25.1C (i.e., most people enter early, although some will continue to enter at later times). If now the number of people remaining in the audience is plotted (Fig. 25.1D), it will be a wider version of
Figure 25.1C, with some people staying until the end, but some others leaving (bored!) before the end. Contrary to Figure 25.1B, Figure 25.1D is no longer a sole measure of the “quality” of the lecture: Its shape reflects also the function describing the audience entrance into the room. To isolate the function solely describing the “quality” of the talk (i.e., Fig. 25.1B, which could only be measured in an ideal situation of an “instantaneous” audience entrance), the contribution of the function describing the audience entrance must be removed. In the perfusion imaging analogy, the theater corresponds to the tissue voxel, the audience corresponds to the contrast agent concentration, the function describing the audience entrance (Fig. 25.1C) corresponds to the AIF, the number of people remaining in the room (Fig. 25.1D) corresponds to the tissue concentration, and the function describing the “quality” of the talk for an ideal “instantaneous” audience (Fig. 25.1B) corresponds to the so-called impulse response function. Finally, the deconvolution in perfusion imaging corresponds to calculating from Figure 25.1D the function describing the “quality” of the talk (Fig. 25.1B) by removing from Figure 25.1D the contribution from the audience entrance (Fig. 25.1C).
Figure 25.1C, with some people staying until the end, but some others leaving (bored!) before the end. Contrary to Figure 25.1B, Figure 25.1D is no longer a sole measure of the “quality” of the lecture: Its shape reflects also the function describing the audience entrance into the room. To isolate the function solely describing the “quality” of the talk (i.e., Fig. 25.1B, which could only be measured in an ideal situation of an “instantaneous” audience entrance), the contribution of the function describing the audience entrance must be removed. In the perfusion imaging analogy, the theater corresponds to the tissue voxel, the audience corresponds to the contrast agent concentration, the function describing the audience entrance (Fig. 25.1C) corresponds to the AIF, the number of people remaining in the room (Fig. 25.1D) corresponds to the tissue concentration, and the function describing the “quality” of the talk for an ideal “instantaneous” audience (Fig. 25.1B) corresponds to the so-called impulse response function. Finally, the deconvolution in perfusion imaging corresponds to calculating from Figure 25.1D the function describing the “quality” of the talk (Fig. 25.1B) by removing from Figure 25.1D the contribution from the audience entrance (Fig. 25.1C).
Geek Box: Residue Function-Derived Formulas for CBF, MTT, and CBV
Originating from indicator dilution theory,1 the following quantitative measures are usually considered in bolus-tracking experiments.
Cerebral Blood Flow (CBF)
CBF represents the amount of tracer that passes through the hemodynamic system per unit time and, given that by definition r(t = 0) = 1 (i.e., following an ideal injection of an infinitesimally narrow bolus, all contrast agent resides within the tissue microvasculature at the beginning [i.e., t = 0]), CBF can be calculated from the initial value of the IRF:
Cerebral Blood Volume (CBV)
CBV represents the total amount of tracer that has passed through the system after administration of the unit narrow impulse injection (i.e., the integral of the IRF):
Equivalence between these two formulations can be found in Straka et al.2
Mean Transit Time (MTT)
MTT is the average time it takes contrast material (and hence blood) to travel through the capillary network in tissue after a single narrow impulse injection. Here, MTT is defined via the central volume principle1 as MTT = CBV/CBF. MTT can be also calculated directly from the residue function, by considering the relationship between the residue function and the probability distribution of transit times in the tissue capillary bed,1 h(t) = −dr(t)/dt (see also Did You Know That? box in left column):
Did you Know That
MTT versus First Moment-Derived Apparent MTT
A commonly used approximation for the tissue MTT, particularly before the widespread use of deconvolution methods, was the first moment of the tissue concentration curve:
This approximation assumes the tissue concentration is an estimate of h(t) (cf. Eq. [8]). But because this is not really the case, MTTapp does not provide accurate estimates of the true MTT. In particular, the CTC first moment is influenced by the AIF, which can vary dramatically between subjects because of the influence of cardiac output, vascular transport, and injection conditions; its use as a measure of perfusion status may be thus potentially misleading (see Fig. 25.2).
Geek Box: Discretization Approximations to the Deconvolution Equation
Because of the relatively coarse sampling interval (typically 1.5 to 2.5 seconds) compared with the bolus first passage (∼10 seconds), discrete deconvolution (even in the absence of noise) results in a distorted CBF·r, and in particular an underestimation of the initial point.1 These discretization errors can be minimized (although not eliminated) by formulating the AIF matrix to assume values of cA(t) and r(t) between measurement points, but without actually subsampling. Assuming linear variation over time, the elements of the AIF matrix become:1
Deconvolution: Ill-Posed Problem and the Need for Regularization
As discussed previously, deconvolution inherently requires a numerical inversion. Therefore, the operation 1/CA(f) (for frequency-domain deconvolution) or the matrix inversion (for deconvolution in time domain) leads to a so-called ill-conditioned deconvolution if the components in question are very small or zero. The problem with the
matrix formulation can be more easily appreciated when the singular value decomposition (SVD)1 is used for solving the matrix inversion. SVD factors the matrix AIF into three matrices: AIF = U.W.V*, where U and V are orthonormal matrices, W is a rectangular diagonal matrix containing the singular values (σij) in decreasing order, and * indicates conjugate transpose. Then, the inverse (for square matrices) or pseudoinverse (for rectangular matrices) of A can be determined as AIF+ = V.W+.U*, where AIF+ is the sought (pseudo)inverse of the matrix AIF, and W+ is the inverse of diagonal matrix W. Per definition, the diagonal matrix W+ can straightforwardly be formed by replacing every diagonal entry σii of W by its reciprocal (i.e., 1/σii) and then by transposing this resulting matrix. The columns of U and V can be interpreted as basis functions (with increasing frequency for decreasing singular value index j): The SVD solution thus corresponds to a decomposition of the residue function in this basis function set. It therefore follows that the contribution from the high-frequency components associated with very small (or zero) σii will be greatly amplified.
matrix formulation can be more easily appreciated when the singular value decomposition (SVD)1 is used for solving the matrix inversion. SVD factors the matrix AIF into three matrices: AIF = U.W.V*, where U and V are orthonormal matrices, W is a rectangular diagonal matrix containing the singular values (σij) in decreasing order, and * indicates conjugate transpose. Then, the inverse (for square matrices) or pseudoinverse (for rectangular matrices) of A can be determined as AIF+ = V.W+.U*, where AIF+ is the sought (pseudo)inverse of the matrix AIF, and W+ is the inverse of diagonal matrix W. Per definition, the diagonal matrix W+ can straightforwardly be formed by replacing every diagonal entry σii of W by its reciprocal (i.e., 1/σii) and then by transposing this resulting matrix. The columns of U and V can be interpreted as basis functions (with increasing frequency for decreasing singular value index j): The SVD solution thus corresponds to a decomposition of the residue function in this basis function set. It therefore follows that the contribution from the high-frequency components associated with very small (or zero) σii will be greatly amplified.
Similarly, computing 1/CA(f) in the frequency-domain deconvolution is not meaningful if the components are very small or zero. In theory, the components of CA(f) or W should be always nonzero because they represent a spectrum of a nonperiodic signal. In practice, however, some of these components can end up being zero (or very close to zero) because of noise. Computation of direct reciprocals of these components will result in infinite or nearly infinite magnitudes of components in the solution, leading ultimately to physically implausible solutions for r(t). More specifically, the components of CA(f) and W that are mostly affected by noise are the high-frequency components, because for typical shapes of the bolus curve, the CTC contains only little power in the high frequencies.
Stabilizing the solution to obtain physically meaningful estimates for r(t), CBF, and mean transit time (MTT) requires treatment of the noise-corrupted components. Most typically, these components are ignored in deconvolution—an operation that is termed regularization. The resulting ř(t) (where ř denotes the regularized version of the solution) is then computed with zero values in locations of W+ or 1/CA(f); that is, in locations where noise-corrupted components were originally present, r is approximated with zero-amplitude basis functions: CBF·ř = AIF+·cT. Therefore, because CBF is estimated from the initial value of the measured IRF (see Eq. [6]), the regularization process [which distorts the shape of r(t)] will have an effect on the estimated perfusion values (see Fig. 25.3). In essence, truncated (or regularized) SVD acts like a low-pass filter. The harsher the regularization filter, the smoother the resulting residue function and thus the more difficult for the measured function to reflect sudden changes in the residue function (e.g., at t = 0). This, in turn, will affect the ability to correctly determine the point in time when the maximum of the residue function will be occurring.
Parametric Deconvolution versus Model-Independent Deconvolution
There are two main approaches to solving the ill-posed deconvolution problem (Eq. [1]): parametric (model-dependent) deconvolution and model-independent deconvolution. In model-dependent deconvolution, the solution to Eq. [1] is constrained by assuming that r(t) is described by a mathematical function, for example, r(t) = exp(−t/MTT), to model a well-mixed compartment.1 In this example, only the free parameter MTT needs to be determined. More complex models can include parameters for arterial delay and dispersion,6,7 or by modeling the microvasculature dynamics (e.g., Mouridsen et al.8). Model-dependent deconvolution has the advantage of ensuring smooth monotonically decreasing estimates of r(t). on the other hand, it can be very sensitive to noise,9 and increasingly so for the flexible models containing a larger
number of free parameters. Furthermore, if the in vivo situation differs from the assumed model, there will be large systematic CBF errors.1 For this reason, model-independent deconvolution approaches are generally preferred and more widely used. In this approach, both CBF and the residue function are considered unknowns (see the fourth Geek Box for a brief description of the most commonly used methods).
number of free parameters. Furthermore, if the in vivo situation differs from the assumed model, there will be large systematic CBF errors.1 For this reason, model-independent deconvolution approaches are generally preferred and more widely used. In this approach, both CBF and the residue function are considered unknowns (see the fourth Geek Box for a brief description of the most commonly used methods).
Geek Box: Common Model-Independent Deconvolution Methods
The key attraction of these techniques is that they do not impose any analytical constraints, or make any assumptions as to the underlying vasculature. As described in Geek Box 1, one well-known model-independent approach is to apply the convolution theorem in the Fourier domain. To stabilize the solution, which depends on 1 over the FT of the AIF, 1/CA(f), frequency components of CA(f) that have a magnitude below a certain threshold are often truncated, that is, appodized or set to zero (see Fig. 25.4). Abrupt truncation of spectra, however, leads to unwanted oscillations (“ringing”) in the time signal. To suppress these oscillations, less abrupt, smoother filters are commonly used:
An alternative approach to the deconvolution in the Fourier domain is based on the algebraic reformulation of Eq. [1] in matrix notation (Eq. [3]). As discussed previously, a very popular approach for time domain–based deconvolution is to invert the system matrix by using SVD. A principal question is then the identification of SVD components that need to be removed. The simplest solution is to choose components below a certain threshold. In early studies,1 the optimal threshold was determined to be 15% to 20% of the maximum singular value (truncated SVD deconvolution, TSVD). The threshold selected in Østergaard et al.1 was a value suitable for typical signal-to-noise ratio (SNR) values in routine DSC-MRI studies. This threshold value, however, can be suboptimal if noise conditions vary (e.g., for MRI sequences with different TE, TR, flip-angle, magnetic field strength, etc.), because the optimal threshold is SNR dependent. Therefore, later approaches aimed to determine the optimal threshold value adaptively. For example, in an attempt to deal with this problem, a look-up table for regularization thresholds based on SNR was proposed.10 Other approaches have also been proposed, which run regularized deconvolution at multiple regularization thresholds and pick the “optimal” threshold based on some property of the solution.3,11 For example, the L-curve criterion,3 one of the simplest and most popular methods, relies on finding the corner of the so-called L-curve. This curve displays the tradeoff between minimizing the residual norm and the constraint norm of the solution (constraint based on a priori properties of the solution, e.g., a nonoscillating smooth residue function). When plotted (for multiple regularization values) on a log-log scale, this curve often has a characteristic L-shaped appearance, with its corner shown to give a regularization parameter that provides an acceptable compromise between data misfit and the regularization term.3 The improvement of these approaches, which result from having to assess multiple regularization values during optimization, comes at the expense of a typical longer computation time.