Temporal resolution and SNR requirements for accurate DCE-MRI measurement of microvascular blood flow and permeability using the AATH model

Introduction: Dynamic contrast-enhanced MRI has been used in conjunction with tracer kinetics modelling in a wide range of tissues for treatment monitoring, oncology drug development and investigation of disease processes. Accurate measurement of model parameters relies on high quality data, and as model complexity increases to allow measurement of more physiological parameters, the data quality requirements are likely to be even more stringent. In this study, the accuracy of the adiabatic approximation to the tissue homogeneity (AATH) model [1] was investigated through simulation. This model allows separation of blood flow from permeability, rather than combining these parameters in a transfer constant (K) as in the more commonly-used Kety model. The effects of temporal resolution and noise levels were examined, and also the impact of an incorrect measurement of the arterial input function (AIF) first pass peak was assessed. Methods: Tissue uptake curves were simulated using the Johnson and Wilson model [2] and the mean AIF from a previous study [3] fitted with a smooth function similar to that used in [4]. Simulated parameters were extraction fraction, E, (0.5–0.9), plasma flow, Fp, (0.03–0.27 ml (ml tissue) min), mean transit time Tc (0.05 1.25 min) and extravascular extracellular volume, ve, (0.1-0.6 ml (ml tissue)), giving a total of 1050 parameter sets. Curves and AIFs were downsampled from the original 1.5 s time resolution (tres) to 3 s, 4.5 s and 6 s, and Gaussiandistributed noise was added with standard deviations of 0.02, 0.05 and 0.1. The curves were fitted using the AATH model as described in [3]. For data simulating a well-designed DCE-MRI experiment (tres = 1.5 s, noise σ = 0.05), curves were fitted using an AIF with the first pass peak height varied between ± 25 % to simulate an incorrectly measured AIF. Results: The influence of temporal resolution and SNR on the fitted AATH parameters is shown in Fig. 1. For each temporal sampling regime and noise level the median percentage errors and interquartile ranges are shown for each parameter, averaged over all parameter values. All parameters were relatively robust to temporal undersampling and noise. Median errors remained below 10 % at 1.5 s temporal sampling across all noise levels, and below 20 % up to 6 s sampling. For Tc, both the median error and the interquartile range increased with noise level and with sampling over 1.5 s. The trend in error with Tc magnitude is plotted in Fig. 2 for high and low quality data, showing very large errors for short Tc. The impact of AIF peak height error on AATH parameters is shown in Fig. 3. Under the conditions of a well-designed DCE-MRI experiment, an underestimate of the AIF peak results in an overestimate of blood flow, and vice-versa. For all other parameters this trend is reversed. Discussion: This simulation study investigated the influence of three key factors on parameter accuracy for the AATH model. The advantage of using this model is that it allows separate estimates of blood flow and permeability, which is not possible with simpler models. Although the AATH model has been used in several previous studies, the expected accuracy of the parameters has remained unclear. The results of this study showed that signal to noise ratio did not have a large influence on parameter accuracy, with errors remaining within ± 10 % for all noise levels at 1.5 s temporal sampling. A temporal resolution of 1.5 s is required if all parameters are to be measured with minimal bias, but this can be relaxed to 6 s if Tc need not be measured and small amounts of bias in E and Fp can be tolerated. Tc showed the largest bias and range of errors and the greatest sensitivity to temporal sampling and noise. This result is consistent with previous studies showing large uncertainties associated with Tc for an individual fit [5] and emphasizes that the sampling interval must be smaller or comparable with Tc [1] to capture information about the initial transit through the vasculature. Fp showed a slight bias when the temporal sampling was inadequate to characterise the initial part of the curve which is most sensitive to this parameter [5]. For ve, slower sampling reduced bias, possibly as a result of parameter covariance with E and Fp. Errors in the measurement of the AIF peak height propagated into errors in parameter estimates, with a 10% error in peak height resulting in an error of at most 10 % in each parameter. The inverse relationship between peak height and blood flow is expected since Fp is the scaling factor for the AIF in the convolution describing the AATH model. Conclusion: Obtaining data with a high temporal resolution is the most critical factor in ensuring accurate parameter estimates using the AATH model, but this requirement can be relaxed if larger biases can be permitted and Tc need not be accurately measured. The trends observed and the data requirements recommended in this study are applicable to implementations of the AATH model in both MRI and CT data. Acknowledgements: We thank Prof D.L. Buckley for providing AIF data. Grant Sponsors: NSERC. References: [1] St Lawrence et al, JCBFM 1998;18:1365 [2] Johnson et al, Am J Physiol 1933;210:1299 [3] Kershaw et al, JMRI 2009;29:641 [4] Parker et al, MRM 2006;56:993 [5] Koh et al, MRM 2006;56:986