Uncertainty estimation in dynamic contrast‐enhanced MRI

Abstract

Using dynamic contrast-enhanced MRI (DCE-MRI), it is possible to estimate pharmacokinetic (PK) parameters that convey information about physiological properties, e.g., in tumors. In DCE-MRI, errors propagate in a nontrivial way to the PK parameters. We propose a method based on multivariate linear error propagation to calculate uncertainty maps for the PK parameters. Uncertainties in the PK parameters were investigated for the modified Kety model. The method was evaluated with Monte Carlo simulations and exemplified with in vivo brain tumor data. PK parameter uncertainties due to noise in dynamic data were accurately estimated. Noise with standard deviation up to 15% in the baseline signal and the baseline T1 map gave estimated uncertainties in good agreement with the Monte Carlo simulations. Good agreement was also found for up to 15% errors in the arterial input function amplitude. The method was less accurate for errors in the bolus arrival time with disagreements of 23%, 32%, and 29% for K(trans) , ve , and vp , respectively, when the standard deviation of the bolus arrival time error was 5.3 s. In conclusion, the proposed method provides efficient means for calculation of uncertainty maps, and it was applicable to a wide range of sources of uncertainty.

Simulating the effect of input errors on the accuracy of Tofts’ pharmacokinetic model parameters

“…Very similar variations of K trans and v e due to errors in T 1 were observed by Di Giovanni et al The same dependences of the calculated K trans and v e on the native K nat trans and v e _ nat were also observed in the same publication. The dependence of K trans and v e on the amplitude of the AIF matches the results published by Garpebring et al [30]. The systematic overestimation of K trans found with the TM applied to data with v p N 0 was also observed in earlier studies [8,44].…”

“…(8)). This approach differs from other studies, where the simulated data were created using methods such as MMID4 or Montecarlo simulations [10,30], and where the error propagation was addressed analytically [24,30]. Though our approach might be biased, the reason behind this choice was that this would make it apparent that a 0% error on any of the IP should result in calculated PMPs without error.…”

“…Orton et al [37] and Mc Grath et al [43] investigated the effect of different forms of the AIF on the results of the PKM. Uncertainty due to noise, inaccurate native T 1 values, bolus arrival time and AIF amplitude were investigated by Garpebring et al [30]. Table 1).…”

“…Previous work has already been published in which the propagation of errors in PKM was addressed [9,[23][24][25][26][27][28][29][30]. These studies either concentrate on the error generated on the PMPs by measurement limitations such as noise [24] or insufficient temporal sampling [26,27], or focus on only one of the two PKM inputs (either AIF [9] or tissue concentration [24,25,28]).…”

State-of-the-art MRI techniques in neuroradiology: principles, pitfalls, and clinical applications

“…An important requirement for the pharmacokinetic modeling is an accurate arterial input function (AIF) ideally derived from the same MRI data [86]. In practice, this is often a challenge as a result either of the location of the target lesion or the measurement method used to acquire the AIF.…”