- Original research
- Open Access
A gamma-distribution convolution model of ^{99m}Tc-MIBI thyroid time-activity curves
- Carl A. Wesolowski^{1, 2}Email authorView ORCID ID profile,
- Surajith N. Wanasundara^{1},
- Michal J. Wesolowski^{1},
- Belkis Erbas^{3} and
- Paul S. Babyn^{1}
- Received: 10 March 2016
- Accepted: 22 November 2016
- Published: 16 December 2016
Abstract
Background
The convolution approach to thyroid time-activity curve (TAC) data fitting with a gamma distribution convolution (GDC) TAC model following bolus intravenous injection is presented and applied to ^{99m}Tc-MIBI data. The GDC model is a convolution of two gamma distribution functions that simultaneously models the distribution and washout kinetics of the radiotracer.
The GDC model was fitted to thyroid region of interest (ROI) TAC data from 1 min per frame ^{99m}Tc-MIBI image series for 90 min; GDC models were generated for three patients having left and right thyroid lobe and total thyroid ROIs, and were contrasted with washout-only models, i.e., less complete models. GDC model accuracy was tested using 10 Monte Carlo simulations for each clinical ROI.
Results
The nine clinical GDC models, obtained from least counting error of counting, exhibited corrected (for 6 parameters) fit errors ranging from 0.998% to 1.82%. The range of all thyroid mean residence times (MRTs) was 212 to 699 min, which from noise injected simulations of each case had an average coefficient of variation of 0.7% and a not statistically significant accuracy error of 0.5% (p = 0.5, 2-sample paired t test). The slowest MRT value (699 min) was from a single thyroid lobe with a tissue diagnosed parathyroid adenoma also seen on scanning as retained marker. The two total thyroid ROIs without substantial pathology had MRT values of 278 and 350 min overlapping a published ^{99m}Tc-MIBI thyroid MRT value. One combined value and four unrelated washout-only models were tested and exhibited R-squared values for MRT with the GDC, i.e., a more complete concentration model, ranging from 0.0183 to 0.9395.
Conclusions
The GDC models had a small enough TAC noise-image misregistration (0.8%) that they have a plausible use as simulations of thyroid activity for querying performance of other models such as washout models, for altered ROI size, noise, administered dose, and image framing rates. Indeed, of the four washout-only models tested, no single model approached the apparent accuracy of the GDC model using only 90 min of data. Ninety minutes is a long gamma-camera acquisition time for a patient, but a short a time for most kinetic models. Consequently, the results should be regarded as preliminary.
Keywords
- Gamma distribution convolution
- ^{99m}Tc-hexakis-methoxy-isobutyl-isonitrile
- Thyroid
- Time-activity curve
- Gamma camera
Background
The use of technetium-99m hexakis-methoxy-isobutyl-isonitrile (^{99m}Tc-MIBI) as a thyroid and parathyroid imaging agent was first proposed in the late 1980s [1]. Since its inception, ^{99m}Tc-MIBI scintigraphy has been shown to be more accurate and sensitive than comparable imaging techniques and is currently used to detect and localize differentiated thyroid cancer [2, 3], parathyroid adenoma [1, 4], parathyroid hyperplasia [5, 6], and parathyroid carcinoma [4, 7].
Several ^{99m}Tc-MIBI image acquisition protocols and analysis techniques have been devised over the last quarter century to aid in the detection of abnormal thyroid and parathyroid tissue, these include dual-phase scintigraphy [8], factor analysis of a dynamic scan [9], and time-activity curve (TAC) analysis of a dynamic scan [10]. TAC analysis provides a means of differentiating between normal and abnormal tissues by comparing of radiotracer uptake and washout in different regions of interest (ROIs). ^{99m}Tc-MIBI does not undergo significant chemical changes in the body and therefore is passively taken up in the thyroid and parathyroid tissues and is cleared rapidly from the blood [11]. TAC analysis also has the potential to provide detailed pharmacokinetic information from ROIs. Pharmacokinetic parameters, such as the tracer mean residence time (MRT) and elimination half-life can be used to quantify and characterize disease states, which ultimately may aid in the diagnosis of thyroid and parathyroid disease.
It is a common practice to inject a bolus of a drug into a vein, but to take images or venous samples elsewhere. Before the time of first arrival of activity in the sampling region, there is no drug and no signal in that remote site. Most pharmacokinetic bolus models ignore this initial mixing and are washout-only models having a maximum amplitude initially. Thus, at the earliest times following a bolus, there is a severe mismatch between a washout model and the signal it models. Consequently, most of the routinely used pharmacokinetic models, e.g., classic compartmental washout models, do a poor job of fitting early-time organ activity [12, 13]. Thus, washout functions do not fit concentration or time-activity curves well enough to serve as accurate simulation study models. To model the entire TAC accurately, it is necessary to use a mathematical model that includes at least some of the drug arrival and organ drug distribution effects. Convolutions of two functions as an impulse response model can approximate the entire TAC accurately [14, 15]. Of these two functions, the first or fast function can be thought of as an organ feed or stimulus function consisting of rapidly changing blood pool background within the organ. The second can be thought of as the response or result of bathing the parenchyma of an organ with drug [16, 17].
While there has been some work on the biodistribution and pharmacokinetics of ^{99m}Tc-MIBI [18, 19], to the best of our knowledge, there has been no presentation of a gamma distribution convolution (GDC) model application in the nuclear or pharmacokinetic literature. GDC models have been applied elsewhere for ecological water storage I/O, waiting times in queueing theory, and in the evaluation of aggregate economic risk of portfolios [20]. Another possibility would be to construct a convolution model from observed input and organ functions using multiple ROIs as has been done for the kidney [16, 21, 22]. However, the advantages of finding a closed form convolution solution for TAC use, like the GDC model, are that only a single organ ROI is used, and that long-term data collection, as needed for washout-only modelling, becomes a shorter term requirement. Finally, unlike convolution methods using observed input functions, a closed form convolution is self-contained and thus useful as a surrogate for otherwise impossible to perform exhaustive testing while still maintaining a degree of realism.
In this context, we aim to demonstrate that following a peripheral intravenous bolus injection, the thyroid has a TAC that can be modelled accurately using the closed form convolution of two gamma distribution (GD) functions. The first GD of the GDC approximates thyroid gland marker arrival and distribution, and the second GD, the thyroid marker washout. We apply this model to TAC’s from dynamic ^{99m}Tc-MIBI scans, and show how it can be used to calculate organ statistics, especially MRT values. Next, we perform a surrogate test example, Monte Carlo simulation of self-consistency of the GDC algorithm to check the accuracy of modelling. Finally, we compare the GDC, a more complete model of concentration, to four of the best available washout-only models for our 90-minute data.
Theory
The amount of daughter (or time activity) at time zero, i.e., the initial EDC functional height, is zero in this particular non-negative-valued Bateman equation. However, when taken out of the parent-daughter decay context, Eq. (1) becomes only approximate. Generally, whatever an exponential distribution (ED) can model is typically better modelled by a gamma distribution (GD) [26]. Exponential distributions are not the preferred shapes to explain organ bolus input function shapes, for which GDs or other functions are more useful [27–29]. To model the prolonged washout kinetics, which models are only fit following peak organ activity, GDs, one of several possible generalizations of EDs, have better fidelity than mono- and bi-exponential models [12, 13, 30–33]. One reason for this is that GD washout models imply zero initial drug volume with no initial mixing [33]. Thus, a GD convolution (GDC) should better fit an organ TAC than Eq. (1), as both the organ delivery and organ washout are more realistically modelled using GDs than EDs. The b = β solution to Bateman Eq. (1) can be written as a gamma density; GD(2, b; t) and its n ^{th} convolution as GD(n, b; t). However, b = β is statistically implausible, and n is only ever an integer, thus despite claims to the contrary [34], GDs do not derive from serial EDCs. Rather, Eq. (1) is a special case GD convolution–see the text following Eq. (5), and a derivation of the GD appears elsewhere [35].
However, tracer residence within the thyroid starts at the time of tracer first arrival within that organ, such that t _{A} is unrelated to thyroid residence time. Moreover, \( {\mathrm{MRT}}_{\mathrm{fast}}=\frac{\mathrm{a}}{\mathrm{b}} \) quantifies the delivery and uptake of tracer in the thyroid, and only \( \mathrm{M}\mathrm{R}{\mathrm{T}}_{\mathrm{WO}}=\frac{\alpha }{\beta } \) relates to how the thyroid washes out of tracer activity once that activity is distributed within the thyroid. That is, MRT_{WO} characterizes the response of thyroid tissue to an instantaneous signal.
Methods
Patient population
The data from three patients were acquired at the Hacettepe University Medical School and were processed blindly at the University of Saskatchewan. The ^{99m}Tc dosages were 740 MBq. Patients 1 and 3 underwent ^{99m}Tc-MIBI scanning for metastatic screening, and were negative. Patient 2 underwent parathyroid scanning for chemical hyperparathyroidism. On scanning, there was ^{99m}Tc-MIBI retention in the lower part of the right thyroid lobe ROI found on histological examination to be a parathyroid adenoma measuring 11 × 11 mm. The GDC curve analysis was performed blindly and without knowledge of the region drawing, the scanned images, or the clinical presentation and results. Clinical correlation was performed only after the quantitative GDC results were obtained.
Data processing
Following a peripheral intravenous bolus injection, 89 one-minute per frame dynamic ^{99m}Tc-MIBI gamma camera images were obtained. Regions of interest (ROIs) were drawn over the right and left thyroid lobe in each frame, and the regional counts per minute were then used to construct time-activity curves (TACs). The TACs were then least Poisson noise model fit with the decay simulated GDC model, Eq. (5)—see the Appendix section—in a separate institution using Mathematica 10.3 programs using an algorithm stop condition of a 10^{-20} relative difference between successive iterations. Caution was taken to find global minima by including reasonable initial parameter range, starting values, but there is no absolute guarantee that global minima will be found with the Nelder-Mead (simplex) regression used. Poisson model noise calculations, regression fitting for Poisson models, corrected fit error quantification, and an algorithm for finding TAC starting-time values, t _{A}, can be found in the Appendix section.
Time-activity curve simulations
Monte Carlo simulations were performed for the purpose of estimating how accurate and precise the GDC results were. The Monte Carlo simulations of each case without decay correction were performed using each case’s GDC parameters as starting values to generate TACs. The TACs were then randomized ×10 by injecting pseudorandom Poisson noise and refit with GDC models using estimated Poisson loss functions to recover the altered (test) parameters. Unlike for the clinical data, the simulations had no injected misregistration. Radioactive decay was not simulated in order to (1) partly compensate for the lack of injected misregistration by increasing the noise to larger levels due to generating Poisson model noise from larger, not decay-diminished signals (2) crosscheck the accuracy of the decay-corrected GDC model used for the clinical cases, (3) simulate drug assay when decay is not a factor, and (4) provide an example of how simulations studies can be used to model variations of experimental conditions and validate techniques.
Washout model analysis with comparison to GDC models
Comparisons of GDC MRT_{WO} results were made to four washout models; two different regression methods of fitting biexponentials, and two different regression methods of fitting single gamma variates. One of these latter, Tk-GV; the Tikhonov regularized gamma variate herein adaptively minimizes the relative error of β or alternatively the minimum relative error of plasma clearance [12, 13, 32]. Briefly, Tk-GV is an approximate inverse solution to Eq. (4). That is for C(t) = GD_{fast}Tk-GV_{WO}, then Tk-GV_{WO} = GD_{fast} ^{(−1)} ⊗ C(t), and Tk-GV fits an inverse (virtual) function, not concentration itself, C(t).
Results
Parameters for the GDC fit model
Pt thyroid^{a} | t _{A} (min) | a | b (min^{−1}) | α | β (min^{−1}) | S (10^{6} cts)^{b} | Fit err %^{c} | Noise % | 1 − R^{2} (%)^{d} |
---|---|---|---|---|---|---|---|---|---|
1L | 0.195 | 0.846 | 0.909 | 0.8577 | 0.001849 | 3.529 | 1.713 | 1.174 | 0.534 |
1R | 0.334 | 3.733 | 5.211 | 0.8661 | 0.004081 | 2.319 | 1.914 | 1.094 | 0.350 |
1T | 0.277 | 1.631 | 2.060 | 0.8645 | 0.003106 | 5.222 | 1.421 | 0.800 | 0.271 |
2L | 0.340 | 1.351 | 1.692 | 0.9408 | 0.004445 | 2.027 | 1.690 | 1.183 | 0.630 |
2R | 0.282 | 0.995 | 0.928 | 0.8668 | 0.001240 | 6.510 | 1.353 | 1.022 | 0.395 |
2T | 0.307 | 1.127 | 1.203 | 0.8998 | 0.002652 | 6.625 | 1.019 | 0.773 | 0.239 |
3L | 0.306 | 7.878 | 11.644 | 0.8918 | 0.002029 | 3.201 | 1.801 | 1.225 | 0.818 |
3R | 0.425 | 3.082 | 4.520 | 0.8942 | 0.002863 | 3.296 | 1.026 | 1.060 | 0.189 |
3T | 0.374 | 6.470 | 9.842 | 0.8947 | 0.002554 | 6.287 | 0.988 | 0.801 | 0.224 |
Decay-corrected MRT values in min for thyroid delivery and washout from 9 GDC models
Patient | MRT_{fast} | MRT_{WO} |
---|---|---|
Thyroid | Delivery | Washout |
a/b | α/β | |
1L | 0.931 | 463.9 |
1R | 0.716 | 212.2 |
1T | 0.792 | 278.3 |
2L | 0.798 | 211.6 |
2R | 1.073 | 699.0 |
2T | 0.937 | 339.3 |
3L | 0.677 | 439.5 |
3R | 0.682 | 312.4 |
3T | 0.657 | 350.3 |
Poisson noise simulations and accuracy of recovery of generating parameters following regression
Parameters | t _{A} | a | b | α | β | S ^{a} | MRT_{WO} |
---|---|---|---|---|---|---|---|
Units | min | none | min^{−1} | none | min^{−1} | 10^{6} counts | min |
Clinical mean values | 0.316 | 3.013 | 4.223 | 0.8863 | 0.00276 | 4.335 | 367.4 |
Simulation mean values^{b} | 0.316 | 3.314 | 4.588 | 0.8859 | 0.00275 | 4.342 | 369.1 |
Units | Percentage (%) | ||||||
Mean simulation CV error^{c} | 2.4 | 0.23 | 1.6 | 0.26 | 3.1 | 2.6 | 0.69 |
Absolute error in percent^{d} | 0.035 | 10.0 | 8.6 | −0.038 | −0.38 | 0.15 | 0.46 |
Units | Probability | ||||||
Probability of no difference^{e} | 0.98 | 0.44 | 0.52 | 0.32 | 0.38 | 0.73 | 0.51 |
As can be seen in Table 4, the smallest and only statistically not significant R^{2} value (0.0183, p = 0.7) was between OLS E2 and GDC. The parathyroid adenoma (2R) was not identified by its MRT_{WO} using OLS E2 and WLS E2, but was the longest MRT_{WO} for all of the GV-based models and for the GDC model. Although the GV washout models; OLS GV, Tk-GV, and the average of OLS GV and Tk-GV were superior to E2 models for the single-case lesion detection, the highest R^{2} value with the GDC of 0.9395 for those washout models occurred for the average MRT_{WO} values of OLS GV and Tk-GV. This averaged model MRT_{WO} value corresponded to a 10.8% coefficient of variation of root mean square error, CV(rmse), agreement with GDC MRT_{WO}. This compares to the Monte Carlo testing of the GDC MRT_{WO}, with a 0.69% (CV) for precision and 0.46% for accuracy—see Table 3.
For the results listed, we processed the TACs as supplied. These neither represented the values for single thyroid lobes nor for whole thyroid glands but rather a mixture of those values. For GDC models of whole thyroid glands listed in Table 4, the IQR of MRT values was 72 min. The GDC IQR from the given mixture of regions was 206 min, and for thyroid lobes without whole thyroid MRT values, the IQR value was 311 min (via Excel PERCENTILE.EXC subtraction or equivalent algorithm).
Discussion
The quality of the fit results illustrated in Fig. 2 and listed in Table 1 suggest that gamma distribution convolutions, GCDs, can be used to form scaled models that accurately follow TACs of ^{99m}Tc-MIBI activity in the thyroid. A GDC implies two gamma distributions, GDs: a GD_{fast} delivery to the thyroid function and a slower GD_{WO} washout from the thyroid function. Note that these separate gamma distributions were not observed directly, but rather were implied by the thyroid ROI data and the GDC fitting parameters. To illustrate these behaviours, the individual GD_{fast}, GD_{WO}, and GDC density functions are displayed in Fig. 1d. That figure shows that the effect of the GD_{fast} on the GDC dissipated in time, so that eventually the GD_{WO} curve converges to the path of the data (after ~1 hr). This, and the results in Table 4, confirm that washout-only models are only indirectly related to TAC or concentration models [14, 15], and that, in turn explains why washout models appear to require more extended time-sampling than the GDC model. For example, the GDC residuals were homoscedastic over the entire sampling space, which is never the case for washout-only models as washout models are at a maximum value initially, when actual concentration is zero. It has been shown in other systems that washout models, even with a 10 min start time, statistically significantly mismatched early data and agreed better with late data [12].
Most pharmacokinetic models are thought of as successful when their curve fit errors average 10% or less [as rmse in 38]. More important than curve fitting error is how precise the models are for the parameter of interest, or in our case, MRT_{WO} . We note that no washout model for predicting GDC MRT_{WO} had less than a 10% precision error measured as CV(rmse)—see Table 4—to unequivocally suggest its use with only 90 min of ^{99m}Tc-MIBI data. The average MRT_{WO} for OLS GV and Tk-GV correlated best to GDC MRT_{WO}. The OLS GV is a direct fit to concentration and Tk-GV is an inverse solution, virtual concentration fit, and which methods have errors in opposite directions when applied to only 90 min of data. This combined MRT measurement had an error of 10.8% compared to the GDC model, i.e., it was almost good enough for clinical use. Thus, the GDC model may find use as a standard for investigating simpler washout models. Herein, each OLS model (OLS E2 and OLS GV) was outperformed for MRT_{WO} R^{2} with GDC by its corresponding weighted alternative; WLS E2 and Tk-GV. Alone amongst the methods tested, MRT_{WO} OLS E2 did not (statistically significantly) regress or correlate to MRT_{WO} GDC (or to Tk-GV). Better correlated was WLS E2. The 1/C(t)^{2} least squares weighting fits biexponential washout models favour fitting the later data to obtain more accurate half-lives [39]. Although some researchers have, without much discussion, used WLS E2 rather than OLS E2 in their patient series [13, 40, 41], direct numerical evidence of improved correlation against better or more complete models, such as that in Table 4, is currently limited. From the Monte Carlo testing—see Table 3—the GDC MRT_{WO} had a 0.69% (CV) precision and 0.46% accuracy, so that comparison to GDC as a measure is not obviously error prone.
No area normalized background subtraction using a second ROI outside of the thyroid ROI was performed for our analysis. For the current work, background was considered to be the blood pooling within the thyroid region as modelled by the fast or delivery function activity, whose short time-duration (MRT_{fast} in Table 2) agrees in a general way with the short time that ^{99m}Tc-MIBI is known to remain in the blood [11]. The background adjustment herein was to treat the blood pooling in the thyroid ROI as the delivery vehicle to the extravascular tissue within the same ROI for later washout. This is similar to the fast decaying quasi-exponential functional treatment of background for ^{99m}Tc-MAG3 renal scintigraphy by the recently developed blood-pool compensation and modified Patlak-Rutland methods both of which do not use classical background regions [17, 42–44]. The shape of the fast curve, GD_{fast}, i.e., the individual 'a' values in Table 1 were variable and implied shapes sometimes heavier tailed than exponential distributions (a < 1), and sometimes lighter-tailed, i.e., closer to normal distributions (a > 1). It would be remiss to imply that the GD_{fast} curve shape is accurately captured, i.e., the GD_{fast} shape coefficient, 'a' was quite variable—see Table 1. The difficulty in quantifying the GD_{fast} shape is likely due to the unavoidable sparseness of nuclear decay data. Using faster framing rates to capture the shape of the fast function better would likely increase both misregistration and noise as these measures covaried (r = 0.77 see Misregistration in the Appendix section). However, the mean MRT_{fast} times from patient studies, numerically a/b, were quite stable as the 'a' and b values were highly correlated (r = 0.99934). From the simulations, the mean standard deviation MRT_{fast} time was only 1.0 s (within cases), where the population standard deviation (between cases) was 0.137 min (8.2 s) suggesting that two parameter GD_{fast} functions are adequate for calculating MRT_{fast} values. The washout function, GD_{WO}, has a shape parameter, α < 1, which yields a curve that is monotonically decreasing (washing out), with very stable shapes herein [CV(α) = 2.9% from patient studies, 2.6% from simulations between cases, and 0.26% from simulations within cases]. The rate scalar (β) of the GD_{WO} is much smaller (i.e., slower) than the rate scalar (b) of the fast curve. Combined with concentration curves, washout models are used for predicting clearance, as they conserve mass. However, washout models do not form accurate models of early time TACs—see Fig. 1d—and [45], and that is a good motive for creating more complete models, i.e., that better fit the data, with the alternatives being (1) to ignore the early data, extend the data collection for hours, and to use models whose misregistration magnifies substantially outside of fit range—see [12, 45], or (2) include the early data and suffer significant, large misregistration within the fit range—see [12]. Despite the pleomorphic shapes of an assumed fast function, and the noise problem, the resulting GDC functions’ misregistration of the TACs counts was only 0.80% with a total fit error of 1.44% (see Misregistration in the Appendix section). Using our threshold measure of goodness-of-fit for curve fit errors of 10%, the GDC TAC error was an order of magnitude better fitting than most models [38].
MRT values were obtained here without performing absolute uptake calculations, drawing more than one ROI per thyroid lobe, or acquiring data for a long time, e.g., 6 h, see [18]. The GDCs models’ decay corrected mean residence times (MRT_{WO}) of ^{99m}Tc-MIBI in the thyroid TACs ranged from 211.6 to 699.0 min. However, not all of our results were from region drawing over normal tissue. We observed considerable differences between the left and right thyroid lobe MRTs, even in our sample of only three cases. Patients 1 and 3 had neoplastic disease elsewhere and no thyroidal metastatic disease detected on scanning. However, patient 2 (2R) had an inferior pole of right lobe of thyroid region 11 × 11 mm parathyroid adenoma. In diseases such as hyperparathyroidism and parathyroid adenoma/carcinoma, thyroid metabolism is presumed normal while the parathyroid tissue is abnormal and typically hyperplastic hypometabolic (containing enlarged, slow to washout parathyroid glands). Indeed, the MRT for that entire right lobe was the slowest in this series at 699 min even though the adenoma was much smaller than the whole ROI. When the patient with thyroid pathology was excluded, the two remaining total thyroid MRTs had values of 278 and 350 min, or not dissimilar to the published thyroid MRT of 314 min obtained using a monoexponential washout model and 6 h of imaging data [18]. This blinded study may have unintentionally detected an adenoma using the GV models. All of the GV models had their longest values associated with the parathyroid adenoma containing thyroid lobe TAC. Usually far outliers (>3 IQR) should receive some comment, and this was the case for OLS GV for the parathyroid adenoma containing thyroid lobe. That GDCs MRTs are proper measurements is evidenced by (1) the adenoma lobe’s MRT being the longest GDC-MRT time, (2) having lowest total (T) thyroid MRT value interpolation error between L and R thyroid lobes,—see Table 4—and (3) the very low misregistration of the GDC model. However, as the interpatient variation of hot spot MRT values is unknown this case series is far too small to the established clinical utility for any MRT calculation. We also perceive a need to establish robustness of the GDC fit algorithm in the clinical setting and to establish any potential applicability to myriad other thyroid pathologies for which the visual clues on scanning are not as obvious as for hot spot imaging. On the other hand, Table 4 shows that the biexponential, OLS E2 and WLS E2, values (1) failed to identify the adenoma containing thyroid lobe, (2) had poor quality interpolation results of L and R thyroid lobe MRTs, and (3) had large modelling errors. Thus, this case series was large enough to demonstrate that the OLS E2 and WLS E2 models can be problematic for identifying hot spots with only 90 min of data. Our hypothetical explanation for this failure is that the MRT_{WO} values were significantly, spuriously, and similarly shortened due to the instant-mixing property of biexponential models. Instant-mixing range-restricts quantifying redistribution, and biexponential models borrow from the physical redistribution to make inflated mass elimination estimates [33]. Moreover, redistribution is the predominate cause of concentration dilution at early times following bolus delivery [33]. We hypothesize that physically, redistribution is generalized and similar for L, R, and T and that redistribution contamination makes the biexponential MRT values too fast and too similar, and thereby not properly interpolative and consequently non-diagnostic.
In prior works, the kinetic behaviour of ^{99m}Tc-MIBI in the thyroid and parathyroid glands has been extensively explored using semi-quantitative parameters or visual examination of images acquired at different times [8–10]. Although severely limited by the retrospective nature of the current study, the GDC model should nonetheless allow for a prospective precise quantitative assessment of disease states using only 90 min of data for pharmacokinetic parameters such as the MRT, or at very least provides a model of concentration/counting for simulation studies for the evaluation of washout models of MRT. This case series was large enough to demonstrate that GDC functions can be of sufficiently high fidelity of fit to imply the use of GDC functions as templates for pharmacokinetic simulation studies, and in that capacity one can consider the results as demonstrating a proof of principle for simulation studies. Indeed, we checked the self-consistency of the clinical decay correction algorithm by comparing its results to the results from simulated data without drug decay.
For washout-only models, one should wait for at least several minutes before starting the fit of polyexponential washout models, and hours before fitting monoexponential washout models, and the fitted data should be collected for as many hours as possible. One alternative to such a fitting strategy is not to attempt to fit concentration itself, but to fit a different regression target, for example, minimum relative error of the rate scalar (β), or weighting toward the later time samples. However, we did not have enough elapsed time for sampling to do this accurately—see the Results section and Table 4. Our other strategy was to use the GDC; a more complete model from the addition of random variables, where chained random variables add by convolution. The use of convolution to create concentration models inclusive of early and late data is hardly new [15]. However, this has usually been done using sums of exponential terms for the washout models, which latter are less accurate and precise washout models than gamma distribution models see Table 4 and [13, 31, 46]. The current work is possibly a first application of a GDC radioactivity (or concentration) model. Many drugs follow GD washout curves [31]. GD washout models extrapolate to late time better than biexponential functions [12], and in their adaptively obtained form from Tikhonov regularization, as Tk-GV, are less affected by altered body fluid states than exponential models [32, 47, 48]. Thus, GDC model simulations with altered count rates or injected noise levels could be used to inspect the effects of altered ^{99m}Tc-MIBI dosage, ROI size, framing rates, or start times for fitting models and regression targets including those for various washout functions and the GDC models themselves.
The Table 3 summary of Monte Carlo noise simulations illustrates the parameter accuracy one can expect from GDC modelling. The Appendix section Eq. (12) was used to estimate start times and appears to have functioned properly with very little error on simulation. The GD_{fast} parameters 'a' and b were somewhat inaccurately recovered following simulation. However, the most important MRT_{WO} results proved to be both accurate (0.46% mean error) and precise (mean CV 0.69%). An accurately determined parameter was α with a within case accuracy error of −0.038%. This parameter determines the shape of the washout curve, which is the predominant shape determinant of the GDC model. This explains the relative superiority of GD washout models to sums of exponential terms, as the latter have a rigid assumed shape and lack a shape parameter. Thus, it should also be possible to use the GDC model to test for the optimal time to first fit a washout model to a TAC, which given the disagreement between early activity and the typical washout model functional height, is generally thought to be ≥ 5 min for gamma variate washout model adaptive fitting or polyexponential model fitting [12, 13, 32] and > 1 h to 6 h, for monoexponential washout model fitting, which latter cannot be directly compared to the time limited gamma camera data herein [18]. In summary, what sample times are needed for fitting a washout model to data so that the parameters obtained are accurate has not been systematically examined, but the methods herein provide a model that should be useable for that purpose.
One of the limitations of the current retrospective study was that the gamma camera acquisitions in the parent series were started after tracer first arrival in the thyroid. However, for the studies processed and for our simulations, a minor delay in the acquisition start time proved to be somewhat forgiving—see the Appendix section for details. Nevertheless, further work on calculating the time of first arrival of activity in the target organ and examination of a larger patient series may be useful. The GDC model was processed using 90 min of data, and although the GDC model MRT_{WO} results appeared to be meaningful thyroid/parathyroid measurements, the GDC model as applied to longer imaging times remains untested. Thus, the conclusions are preliminary.
Conclusions
By using thyroid ^{99m}Tc-MIBI scintigraphy data from 3 patients, we generated 9 TACs and 90 Monte Carlo TAC simulations using gamma distribution convolution (GDC) models. The GDC models fit the TAC data with high fidelity and small enough TAC noise-image misregistration (0.8%) that they have a plausible use as simulations of thyroid activity for querying performance of other models, such as washout models, for altered ROI size, noise, administered dose and image framing rates. Monte Carlo accuracy testing results for all of the GDC model parameter values were good with a GDC MRT accuracy of 0.46%, despite fitting only 90 min of data. Since the GDC model is an actual concentration model, it does not have to decay several half-lives in order to obtain model parameters or validate the methodology. The 4 washout-only models applied to the same data all exceeded 10% precision error compared to the GDC, with the apparent excess error suspected to be from having insufficient temporal data for washout modelling.
The fast computation of this relies on \( {}_1\mathrm{F}_1\left(\mathrm{A};\mathrm{B};\mathrm{Z}\right)=\frac{\Gamma \left(\mathrm{B}\right)}{\Gamma \left(\mathrm{B}-\mathrm{A}\right)\Gamma \left(\mathrm{A}\right)}{\displaystyle {\int}_0^1{\mathrm{e}}^{\mathrm{Z}\;t}{u}^{\mathrm{A}-1}{\left(1-u\right)}^{\mathrm{B}-\mathrm{A}-1}\mathrm{d}u} \), where A, B, and Z are variables and where the function is finite for all finite values of those variables. Also, we set GDC(⋅⋅⋅⋅; τ = 0) = 0 to replace the 0^{0}indeterminacy of τ ^{a + α− 1} at a + α = 1, τ = 0 with its limiting value of 0.
Declarations
Acknowledgments
Partial funding for this work was provided from grant J2012-114 of the Sylvia Fedoruk Canadian Centre for Nuclear Innovation and grant 110s1002 of TUBITAK (the Scientific and Technological Research Council of Turkey).
Authors’ contributions
CW provided the theory, processed and analysed data, and wrote the first draft of the manuscript. SW, PB, and MW have participated in computer programming, organizing and interpretation of data, and drafting and editing the manuscript. BE designed the study, obtained ethics approval, collected and provided patient data, and wrote the patient information section. All authors read and approved the final manuscript.
Competing interests
The authors declare that they have no direct competing interests. The Tk-GV implementation mentioned is patented [50].
Ethics approval and consent to participate
The data were obtained in accordance with the laws in effect in Turkey from a Hacettepe University Medical School Ethics Committee approved study (TBK 09/1-29).
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
- Coakley AJ, Kettle AG, Wells CP, O'Doherty MJ, Collins REC. ^{99m}Tc sestamibi - a new agent for parathyroid imaging. Nucl Med Commun. 1989;10:791–4.View ArticlePubMedGoogle Scholar
- Alonso O, Lago G, Mut F, Hermida J, Nunez M, De Palma G, Touya E. Thyroid imaging with Tc-99 m MIBI in patients with solitary cold single nodules on pertechnetate imaging. Clin Nucl Med. 1996;21:363–7.View ArticlePubMedGoogle Scholar
- Mezosi E, Bajnok L, Gyory F, Varga J, Sztojka I, Szabo J, Galuska L, Leovey A, et al. The role of technetium-99 m methoxyisobutylisonitrile scintigraphy in the differential diagnosis of cold thyroid nodules. Eur J Nucl Med Mol Imaging. 1999;26:798–803.View ArticleGoogle Scholar
- Ruda JM, Hollenbeak CS, Stack Jr BC. A systematic review of the diagnosis and treatment of primary hyperparathyroidism from 1995 to 2003. Otolaryngol Head Neck Surg. 2005;132:359–72. doi:10.1016/j.otohns.2004.10.005.View ArticlePubMedGoogle Scholar
- Blocklet D, Martin P, Schoutens A, Verhas M, Hooghe L, Kinnaert P. Presurgical localization of abnormal parathyroid glands using single injection of technetium-99 m methoxyisobutylisonitrile: comparison of different techniques including factor analysis. Eur J Nucl Med Mol Imaging. 1997;24:46–51.View ArticleGoogle Scholar
- Chesser AMS, Carroll MC, Lightowler C, Macdougall IC, Britton KE, Baker LRI. Technetium-99 m methoxy isobutyl isonitrile (MIBI) imaging of the parathyroid glands in patients with renal failure. Nephrol Dial Transpl. 1997;12:97–100.View ArticleGoogle Scholar
- Lu G, Shih W-J, Xiu J-Y. Technetium-99 m MIBI uptake in recurrent parathyroid carcinoma and brown tumors. J Nucl Med. 1995;38:811–3.Google Scholar
- Taillefer R, Boucher Y, Potvin C, Lambert R. Detection and localization of parathyroid adenomas in patients with hyperparathyroidism using a single radionuclide imaging procedure with technetium-99 m-sestamibi (double-phase study). J Nucl Med. 1992;33:1801–7.PubMedGoogle Scholar
- Billotey C, Aurengo A, Najean Y, Sarfati E, Moretti J-L, Toubert M-E, Rain J-D. Identifying abnormal parathyroid glands in the thyroid uptake area using technetium-99 m-Sestamibi and factor analysis of dynamic structures. J Nucl Med. 1994;35:1631–6.PubMedGoogle Scholar
- O'Doherty MJ, Kettle AG, Wells P, Collins R, Coakley AJ. Parathyroid imaging with technetium-99 m-sestamibi: preoperative localization and tissue uptake studies. J Nucl Med. 1992;33:313–8.PubMedGoogle Scholar
- Baggish AL, Boucher CA. Radiopharmaceutical agents for myocardial perfusion imaging. Circulation. 2008;118:1668–74.View ArticlePubMedGoogle Scholar
- Wesolowski CA, Puetter RC, Ling L, Babyn PS. Tikhonov adaptively regularized gamma variate fitting to assess plasma clearance of inert renal markers. J Pharmacokinet Pharmacodyn. 2010;37:435–74. doi:10.1007/s10928-010-9167-z.View ArticlePubMedPubMed CentralGoogle Scholar
- Wanasundara SN, Wesolowski MJ, Barnfield MC, Waller ML, Murray AW, Burniston MT, Babyn PS, Wesolowski CA. Accurate and precise plasma clearance measurement using four 99mTc-DTPA plasma samples over 4 h. Nucl Med Commun. 2016;37:79–86. doi:10.1097/MNM.0000000000000405.PubMedGoogle Scholar
- Rosso L, Brock CS, Gallo JM, Saleem A, Price PM, Turkheimer FE, Aboagye EO. A new model for prediction of drug distribution in tumor and normal tissues: pharmacokinetics of temozolomide in glioma patients. Cancer Res. 2009;69:120–7. doi:10.1158/0008-5472.CAN-08-2356.View ArticlePubMedGoogle Scholar
- Roberts MS, Donaldson JD, Rowland M. Models of hepatic elimination: comparison of stochastic models to describe residence time distributions and to predict the influence of drug distribution, enzyme heterogeneity, and systemic recycling on hepatic elimination. J Pharmacokinet Biopharm. 1988;16:41–83.View ArticlePubMedGoogle Scholar
- Wesolowski CA, Conrad GR, Kirchner PT. A direct modeling approach to the early renal vascular transit of ^{99m}Tc chelates. Med Phys. 1987;14:1032–41. doi:10.1118/1.596117.View ArticlePubMedGoogle Scholar
- Wesolowski MJ, Conrad GR, Šámal M, Watson G, Wanasundara SN, Babyn P, Wesolowski CA. A simple method for determining split renal function from dynamic 99mTc-MAG3 scintigraphic data. Eur J Nucl Med Mol Imaging. 2016;43:550–8. doi:10.1007/s00259-015-3216-1.View ArticlePubMedGoogle Scholar
- Savi A, Gerundini P, Zoli P, Maffioli L, Compierchio A, Colombo F, Matarrese M, Deutsch E. Biodistribution of Tc-99 m methoxy-isobutyl-isonitrile (MIBI) in humans. Eur J Nucl Med. 1989;15:597–600.View ArticlePubMedGoogle Scholar
- Dizdarevic S, Aplin M, Ryan N, Holt S, Goldberg L, Miles KA, Peters M. Renal and hepatic kinetics of Tc-99 m-labelled Heaxakis-methoxy-isobutyl isonitrile. Drug Metab Lett. 2012;6:242–6.View ArticlePubMedGoogle Scholar
- Di Salvo F. The exact distribution of the weighted convolution of two gamma distributions. Atti della XLIII Riunione Scientifica SIS (Acts of the 43 Annual Meeting of the Italian Scientific Society). Università di Torino: Società Italiana di Scientifica; 2006. p. 511-4. http://old.sis-statistica.org/files/pdf/atti/Spontanee%202006_511-514.pdf.
- Conrad GR, Wesolowski CA, Berbaum KS, Jensen KC, Kirchner PT. Renal arteriovenous transit times of technetium-radiolabeled chelates. J Nucl Med. 1987;28:1134–43.PubMedGoogle Scholar
- Chandler ST, Gibson CJ, Elliott L. Models of renal blood flow and their use in the detection of renal artery stenosis. Nucl Med Commun. 1990;11:421–36.View ArticleGoogle Scholar
- Widrow B, Kollar I, Liu M-C. Statistical theory of quantization. IEEE Trans Instrum Meas. 1996;45:353–61.View ArticleGoogle Scholar
- Moschopoulos PG. The distribution of the sum of independent gamma random variables. Ann Inst Stat Math. 1985;37:541–4.View ArticleGoogle Scholar
- Garrett ER. The Bateman function revisited: a critical reevaluation of the quantitative expressions to characterize concentrations in the one compartment body model as a function of time with first-order invasion and first-order elimination. J Pharmacokinet Pharmacodyn. 1994;22:103–28.View ArticleGoogle Scholar
- Stewart T, Strijbosch LWG, Moors JJA, Batenburg PV. A simple approximation to the convolution of gamma distributions (Revision of DP 2006-27). Tilburg University: Center for Economic Research, Discussion Paper 2007-70; 2007.Google Scholar
- Maltz DL, Treves S. Quantitative radionuclide angiocardiography determination of Qp: Qs in children. Circulation. 1973;47:1049–56.View ArticlePubMedGoogle Scholar
- Patil V, Johnson G. An improved model for describing the contrast bolus in perfusion MRI. Med Phys. 2011;38:6380–3.View ArticlePubMedPubMed CentralGoogle Scholar
- Horsfield MA, Thornton JS, Gill A, Jager HR, Priest AN, Morgan B. A functional form for injected MRI Gd-chelate contrast agent concentration incorporating recirculation, extravasation and excretion. Phys Med Biol. 2009;54:2933. doi:10.1088/0031-9155/54/9/023.View ArticlePubMedGoogle Scholar
- Wise ME. The evidence against compartments. Biometrics. 1971;27:262.Google Scholar
- Wise ME. Negative power functions of time in pharmacokinetics and their implications. J Pharmacokinet Pharmacodyn. 1985;13:309–46.View ArticleGoogle Scholar
- Wesolowski CA, Ling L, Xirouchakis E, Burniston MT, Puetter RC, Babyn PS, Giamalis IG, Burroughs AK. Validation of Tikhonov adaptively regularized gamma variate fitting with 24-h plasma clearance in cirrhotic patients with ascites. Eur J Nucl Med Mol Imaging. 2011;38:2247–56. doi:10.1007/s00259-011-1887-9.View ArticlePubMedPubMed CentralGoogle Scholar
- Wesolowski CA, Wesolowski MJ, Babyn PS, Wanasundara SN. Time varying apparent volume of distribution and drug half-lives following intravenous bolus injections. PLoS One. 2016;11, e0158798. doi:10.1371/journal.pone.0158798.View ArticlePubMedPubMed CentralGoogle Scholar
- Davenport R. The derivation of the gamma-variate relationship for tracer dilution curves. J Nucl Med. 1983;24:945–8.PubMedGoogle Scholar
- Singh VP, Singh K. Derivation of the gamma distribution by using the principle of maximum entropy (Pome). J Am Water Resour Assoc. 1985;21:941–52. doi:10.1111/j.1752-1688.1985.tb00189.x.View ArticleGoogle Scholar
- Weisstein EW. Confluent hypergeometric function of the first kind. From MathWorld--A Wolfram Web Resource. 2011. http://mathworld.wolfram.com/ConfluentHypergeometricFunctionoftheFirstKind.html. Retrieved Oct. 14, 2014.
- Arfken G. Confluent Hypergeometric Functions. §13.6 in Mathematical Methods for Physicists, 3rd ed. Orlando: Academic Press; 1985. p. 753-8. ISBN: 978-0-12-059820-5.Google Scholar
- Burger D, Ewings F, Kabamba D, L'homme R, Mulenga V, Kankasa C, Thomason M, Gibb DM, et al. Limited sampling models to predict the pharmacokinetics of nevirapine, stavudine, and lamivudine in HIV-infected children treated with pediatric fixed-dose combination tablets. Ther Drug Monit. 2010;32:369–72.View ArticlePubMedGoogle Scholar
- Toutain P-L, Bousquet-Mélou A. Plasma terminal half-life. J Vet Pharmacol Ther. 2004;27:427–39. doi:10.1111/j.1365-2885.2004.00600.x.View ArticlePubMedGoogle Scholar
- Florijn KW, Barendregt JNM, Lentjes EGWM, van Dam W, Prodjosudjadi W, van Saase JLCM, van Es LA, Chang PC. Glomerular filtration rate measurement by single-shot injection of inulin. Kidney Int. 1994;46:252–9. doi:10.1038/ki.1994.267.View ArticlePubMedGoogle Scholar
- Orlando R, Floreani M, Padrini R, Palatini P. Determination of inulin clearance by bolus intravenous injection in healthy subjects and ascitic patients: equivalence of systemic and renal clearances as glomerular filtration markers. Br J Clin Pharmacol. 1998;46:605–9. doi:10.1046/j.1365-2125.1998.00824.x.View ArticlePubMedPubMed CentralGoogle Scholar
- Blaufox MD. Renal background correction and measurement of split renal function: the challenge. Eur J Nucl Med Mol Imaging. 2016;43:548–9. doi:10.1007/s00259-015-3253-9.View ArticlePubMedGoogle Scholar
- Samal M, Brink A, Wesolowski MJ, Wesolowski CA. Preliminary validation of blood-pool compensation method for assessment of relative renal uptake in children. Eur J Nucl Med Mol Imaging. 2015;42(Supplement 1):S771.Google Scholar
- Conrad GR. How robust are the new blood-pool compensation (BPC) and new modified Patlak-Rutland (MP-R) methods for determining split renal function in ^{99m}Tc-MAG3 scintigraphy? J Nucl Med. 2016;57(Supplement 2):542.Google Scholar
- Wanasundara SN, Wesolowski MJ, Puetter RC, Burniston MT, Xirouchakis E, Giamalis IG, Babyn PS, Wesolowski CA. The early plasma concentration of 51Cr-EDTA in patients with cirrhosis and ascites: a comparison of three models. Nucl Med Commun. 2015;36:392–7. doi:10.1097/MNM.0000000000000255.View ArticlePubMedPubMed CentralGoogle Scholar
- Perkinson AS, Evans CJ, Burniston MT, Smye SW. The effect of improved modelling of plasma clearance in paediatric patients with expanded body spaces on estimation of the glomerular filtration rate. Physiol Meas. 2010;31:183–92. doi:10.1088/0967-3334/31/2/005.View ArticlePubMedGoogle Scholar
- Wickham F, Burniston MT, Xirouchakis E, Theocharidou E, Wesolowski CA, Hilson AJW, Burroughs AK. Development of a modified sampling and calculation method for isotope plasma clearance assessment of the glomerular filtration rate in patients with cirrhosis and ascites. Nucl Med Commun. 2013;34:1124–32. doi:10.1097/MNM.0b013e32836529ab.View ArticlePubMedGoogle Scholar
- Wesolowski CA, Wanasundara SN, Wesolowski MJ, Barnfield MC, Waller ML, Babyn PS. Increased extracellular fluid volume correlates to increased overestimation of 4-h clearance values to a different degree for each of three clearance methods. J Nucl Med. 2016;57(Supplement 2):543.Google Scholar
- Bé M-M, Chisté V, Dulieu C, Browne E, Chechev V, Kuzmenko N, Helmer R, Nichols A, et al. Table of radionuclides (Vol. 1-A = 1 to 150). France: Sèvres: Bureau International des Poids et Mesures; 2004.Google Scholar
- Wesolowski CA, Babyn P, Puetter R. Method for evaluating renal function. United States Patent US 8,738,345 B2 May 27, 2014Google Scholar