An open source software for analysis of dynamic contrast enhanced magnetic resonance images: UMMPerfusion revisited

Perfusion imaging has become an important image based tool to derive the physiological information in various applications, like tumor diagnostics and therapy, stroke, (cardio-) vascular diseases, or functional assessment of organs [1, 2]. Mostly this technique is applied in magnetic resonance imaging (MRI) [3] but also emerges into the field of computed tomography (CT) [4, 5] and ultrasound (US) [6]. Especially in MRI, this technique benefits of good tissue contrast, that it is noninvasive and without the application of ionized radiation. A common approach to measure perfusion using MRI is dynamic contrast enhanced (DCE) MRI, using T1-weighted sequences to record the local signal change due to the contrast agent bolus passing through the observed area. By applying so called pharmacodynamic models to the data hemodynamic parameters like the blood flow (or perfusion), blood volume, mean transit time or the extravasation of the contrast agent from the blood stream e.g., into the interstitial space can be calculated. In recent years, added value of DCE-MR perfusion imaging has been reported in various application, e.g., for kidney [7‐9], liver [10], or heart disease [11]. In prostate cancer DCE-MR perfusion imaging has developed as one part of a multi-parametric approach to stage cancer [12, 13]. It is also applied in preclinical functional imaging [14, 15]. However, even after 20 years of intense research in this field, perfusion imaging by DCE-MRI still remains a research tool without a broad clinical usage. One problem is the lack of standardization in technical aspects which have to be considered for successful quantitative evaluation, including sequence and contrast agent dose optimization [16] model selection [17], correct selection of the arterial input function [18, 19], or correction of motion artifacts [20, 21]. Recently, efforts are made to overcome this, e.g., by the Quantitative Imaging Biomarkers Alliance (QIBA) [22] of the Radiological Society of North America (RSNA) or the EIBALL – European Imaging Biomarkers Alliance [23].

The second problem is a lack of tools that allow a direct integration into the diagnostic workflow in radiology. To the best of our knowledge, apart from the work presented herein, there are only few research tools that are also integrated into a DICOM workstation [24, 25] . Research tools that allow calculations of hemodynamic parameters are developed often as offline solutions and the clinician has to transfer the large image data sets to a separate workstation for analysis [3, 24‐28]. Furthermore, to include the results in the clinical workflow they have to be transferred back into the diagnostic system. However, results obtained from most research software are stored in various formats that could not easily be converted to DICOM objects to be stored in picture archive and communication systems (PACS) [24]. Certainly, the aforementioned procedure is feasible in the research context investigating small patient groups; however, in daily practice this becomes cumbersome.

Commercial software solutions to analyse DCE-MRI data exist and they allow integration into the clinical environment. This comprises products of independent companies but also solutions provided by the vendors of the MR scanners, but suffering of multi-vendor capability. Furthermore, a major problem of these solutions is that they are black-box, which means that validation and absolute benchmarking is difficult. This has real clinical implications, as demonstrated recently in the study by Heye et al. [26] which reported that a “considerable variability for DCE MR imaging pharmacokinetic parameters (Ktrans, kep, ve, iAUGC) was found among commercially available perfusion analysis solutions” and that therefore “clinical comparability across perfusion analysis solutions is currently not warranted”.

In addition, such software is expensive in respect to the cost-benefit ratio: available commercial software solutions are often dedicated to a single application, i.e., heart, brain, or prostate perfusion [27] which does not allow for easy extension and adaption beyond the intended usage in the clinical situation [28]. Therefore, only few licenses or dedicated workstations are usually purchased which prevents ubiquitous usage [29, 30].

Recently our group presented a perfusion analysis tool (UMMPerfusion) that aimed at overcoming some of the aforementioned problems [31]. In the initial version of our software, we provided means for quality assurance by visualizing the arterial input function (AIF) online while drawing its respective region of interest and by generating automatically a report logging all settings of the respective data analysis session. The software itself was designed as a plugin for the Open Source DICOM Workstation OsiriX [32, 33] which can be fully embedded into the radiological workflow [30] and thereby, calculated results by our software, too. The decision to select OsiriX, besides that it has been installed for research and clinincal use in our Radiology department was that OsiriX became a very popular and powerful software with more than 40.000 users worldwide at very low costs. By implementing the perfusion anaylsis software as OsiriX plugin and Open Source, we hope to reach a large number of users and to bring perfusion imaging forward by providing analysis software.

To calculate hemodynamic parameters, however, so far only a model-free deconvolution approach was implemented. Pharmacokinetic models reported in the literature [34] offer additional parameters, e.g., permeability or extravascular extracellular volume and describe the tissue in more detail.

In this paper, we will present recent extensions of our software. This comprises the implementation of several well established compartment models and their integration into the plugin and the quality management developed for this software.

Our software plugin was implemented and tested on Mac OSX systems version 10.8.x using OsiriX versions 5.5.1, 5.6, and 5.9. Apart from these OsiriX versions, the software may work but no tests by the authors were performed so far.

Figure 5 depicts the AIF display and the GUI for an example data analysis of the prostate. The result of this ROI based analysis using the 2CUM is depicted in Fig. 6 while Fig. 7 shows the corresponding report that is created when exporting the data into the OsiriX database. The report is a DICOM object that can be archived with the patient data into PACS systems.

In the ROI based analysis as well as for the report, the calculated parameters derived from fitting the model to the data are listed. Also, the fit itself to the data is visualized in a plot. To assess the quality of the fit not only visually, we provide two goodness-of-fit measures (χ ², AIC). While the χ ² can be used to judge if the model fit was good, the AIC can be used to compare two or more models given the data.

Similar, the results of the other implemented compartment models for a ROI based analysis would look alike. Analyzing the data using the pixel based calculation will result in a parametric map for each of the respective parameters of corresponding compartment model. As an example, Fig. 8 depicts such an analysis for a DCE-MRI data set of the kidney employing the 2CFM.

To evaluate the implemented compartment models we used a procedure previously described in [31] utilizing a reference test data set constructed from a DCE-MRI data set of the prostate. A time series with 100 time points was constructed taking a matrix of 8 × 4 pixels as one slice. Half of these pixels were taken from a vessel representing the AIF in the original data while the remaining pixels were sampled from prostate tissue. To calculate reference values for each compartment model, the software PMI [40], an authorative research tool for perfusion analysis was selected. Obtained reference values for each parameter and the corresponding values of our software were compared. In all settings, no differences between reference and our implementation were detected (see Table 2). By this we conclude that the implementation of the algorithm is technically correct.

In order to process data by our plugin, two prerequisites are required, a) the data has to be in DICOM format and b) it must be loaded into the OsiriX 4D viewer. To further evaluate the robustness of our software, also for processing image data from different vendors, perfusion data sets from the three main vendors of clinical MR systems (Siemens, GE, Philips) were collected and processed successfully. The major challenge in processing data sets of the different vendors is that information of the temporal resolution is stored differently in the DICOM headers, especially for the Philips multi frame storage format (see Table 3). If no such timing information can be extracted from the DICOM header, at the moment, no calculation of the models is possible. We are currently also testing to read DICOM data provided by small animal scanners (Bruker). DCE-MRI data of 2D acquisitions could be successfully analysed.

In addition special slice positioning (e.g., 1 transversal slice, 4 coronal slices) reported in the literature [1, 7] caused problems when loaded into the OsiriX 4D viewer, a prerequisite for our plugin to detect the temporal domain of the data. Using an option in OsiriX to resort this data, the 4D viewer could be opened and thereafter, our plugin could process this data without problems (see Fig. 9).

A current drawback of our implementation is that the non-linear fitting of parametric maps is time consuming. To reduce the computation time, at present a rectangular ROI has to be placed in the DCE-MRI image series. To further improve the computational speed of the calculating parametric maps, a linear least squares approach as proposed by Flouri et al. [36] will be explored. Furthermore, an implementation of the compartment models in a Graphics Processing Unit (GPU) will be considered.

Besides our plugin several research tools exist for perfusion analysis in DCE-MRI [47‐52] which might allow for benchmarking and comparison of the different solutions. In this work, we only compared our algorithms against the reference implementation in PMI [40] to verify the technical correctness. A comparison of PMI against other perfusion analysis software using the QIBA protocol and different levels of noise was reported by Cron et al. [53]. Beuzit et al. compared our plugin to four other software solutions, including commercial software from all three main vendors using simulated and measured data [54]. In this study, the ETM was used and the authors reported a bias for all solutions and pharmacokinetic parameters ranging from 0.19 min⁻¹ to 0.09 min⁻¹ for Ktrans, −0.15 to 0.01 for ve, and −0.65 to 1.66 mmol/L⁻¹/min. In both studies the variances in the parameters between the different software solutions were attributed to various reasons. Cron et al. observed increased unphysiological values with increasing noise while Beuzit et al. stated that probably the (not documented) fitting routine might have had an influence on the results. As stated by Heye et al. such comparison might be in general difficult [26], especially for the commercial solution since little is known about their implementation. Available research tools are implemented on various platforms, requiring different input formats and outputs and eventually implemented different fitting algorithms which make a comparison difficult to interpret. Furthermore, data sets with known true values or available gold standard and fully control on the measured or simulated data should be employed when comparing and validating software to minimize e.g., inter patient variability [55]. All this implies that there is a need in standardizing the DCE-MRI perfusion analysis.

Compared to the above mentioned research tools our plugin underwent an in house certification process. This process which involves risk analysis and thorough documentation also gave insights howto improve the workflow, the structure of the source code and to prevent errors caused accidently by users and thereby improve stability of the plugin. Eventually, it allows for using our plugin for research but also for clinical routine [56]. All documents and procedures of this certification are documented at our website (http://​www.​opossumm.​de) to help others to perform an in house certification by themselves or to adapt the procedure according to their local regulations.

We developed open source software to analyse DCE-MRI perfusion data. The software is designed as plugin for the DICOM Workstation OsiriX. It features a clean GUI and provides a simple workflow for data analysis while it could also be seen as a toolbox providing an implementation of several recent compartment models, adapted from the software PMI, to be applied in research tasks. Integration into the infrastructure of a radiology department is given via OsiriX. Results can be saved automatically and reports generated automatically during data analysis to ensure certain quality control.

Compiled binaries and source code of our Open Source software is available for download via the OpossUMM platform (http://​www.​opossumm.​de). The software requires at least Mac OSX 10.8.x and was tested with OsiriX versions 5.5, 5.6 and 5.9. For optimal performance and to handle large datasets, the installation of the commercial available 64bit extension of OsiriX (now included in OsiriX MD) is suggested.