UMMPerfusion: an Open Source Software Tool Towards Quantitative MRI Perfusion Analysis in Clinical Routine

Abstract

To develop a generic Open Source MRI perfusion analysis tool for quantitative parameter mapping to be used in a clinical workflow and methods for quality management of perfusion data. We implemented a classic, pixel-by-pixel deconvolution approach to quantify T1-weighted contrast-enhanced dynamic MR imaging (DCE-MRI) perfusion data as an OsiriX plug-in. It features parallel computing capabilities and an automated reporting scheme for quality management. Furthermore, by our implementation design, it could be easily extendable to other perfusion algorithms. Obtained results are saved as DICOM objects and directly added to the patient study. The plug-in was evaluated on ten MR perfusion data sets of the prostate and a calibration data set by comparing obtained parametric maps (plasma flow, volume of distribution, and mean transit time) to a widely used reference implementation in IDL. For all data, parametric maps could be calculated and the plug-in worked correctly and stable. On average, a deviation of 0.032 ± 0.02 ml/100 ml/min for the plasma flow, 0.004 ± 0.0007 ml/100 ml for the volume of distribution, and 0.037 ± 0.03 s for the mean transit time between our implementation and a reference implementation was observed. By using computer hardware with eight CPU cores, calculation time could be reduced by a factor of 2.5. We developed successfully an Open Source OsiriX plug-in for T1-DCE-MRI perfusion analysis in a routine quality managed clinical environment. Using model-free deconvolution, it allows for perfusion analysis in various clinical applications. By our plug-in, information about measured physiological processes can be obtained and transferred into clinical practice.

Appendix

Truncated Singular Value decomposition for Perfusion Analysis

For the tracer kinetic analysis, we modified the deconvolution approach proposed by Ostergaard et al. [31]. This is based on the basic relation for residue detection in linear and stationary systems, which relates the tissue concentrations to the concentrations in the arterial plasma by a convolution:

$$ C(t) = F{C_a}(t) \otimes R(t), $$

(1)

with C(t) the measured tissue concentration, C _a (t) the arterial plasma concentrations, R(t) the tissue residue function, and F is the plasma flow. The relation is discretized assuming linear interpolation between measured points, using the Volterra formula in [32]. This produces a matrix equation

$$ {{\bf C}} = \Delta t{{\bf AI}}, $$

(2)

where A is an n × n matrix with n is the number of time points, depending on the values of C _a only. The singular value decomposition

$$ {{\bf A}} = {{\bf US}}{{{\bf V}}^{{{\bf T}}}} $$

(3)

is calculated, all singular values S below a given cut-off λ are set to zero, and the pseudo-inverse

$$ {{{\bf A}}^{{ + }}} = {{\bf V}}{{{\bf S}}^{{ + }}}{{{\bf U}}^{\text{T}}} $$

(4)

is calculated. The impulse response I _k for each pixel k is then calculated by

$$ {I_k} = \Delta {t^{{ - 1}}}{A^{{ + }}}{C_k} $$

(5)

The value of the regularization parameter λ is fixed in this algorithm. The software sets a default value of 0.15 max (S), but allows the user to select a different value if required. A more detailed discussion of selecting the regularization parameter can be found in [30, 54]. To parameterize the results, the plasma flow F _k of each pixel k is calculated as

$$ {F_k} = { \max }\left( {{I_k}} \right) $$

(6)

Using the central-volume theorem and the definition of the residue function we can also determine the volume of distribution as the area under the impulse response

$$ {V_k} = \Delta t\sum\limits_i {{{{\bf I}}_{{ki}}}}, $$

(7)

and the mean transit time

$$ {T_k} = \frac{{{V_k}}}{{{F_k}}} $$

(8)