The training dataset is attached to the supporting online material of this publication. The imaging and biosignal data are raw data, i.e. no segmentation, baseline correction, retrospect filtering or interpolation were performed. Thus it is the choice of the users what methods to consider.
MRI data
The first image stack (Series 301, Additional file 1), the survey, is included for the purpose of providing information on the reference markers, the torso geometry and the heart position within. Figure 1 displays a subset of this image stack. Note the reference markers in IM_0090 and IM_0108, which are essential for the common coordinate system.
Series 601 (Additional files 2 and 3) contains for several heart slices the CINE sequence over the whole heart beat period. The images selected in Figure 6 reflect the time instances indicated by red cursors in subsequent figures.
The last two series (701, Additional file 4 and 801, Additional file 5) are the high resolution datasets at diastole and systole, respectively, which should serve as input data for the heart modelling algorithms.
Unfortunately, the raw image data are not the data the modeler wants to know. What is really needed are the volumetric spatial coordinates of the boundaries between different tissue or body compartments, e.g. between blood volume and endocardium, etc. In other words, a thorough segmentation of the image data is an essential processing step to providing the geometrical input data relevant to the model. It is intentional that this segmentation has not been applied to the data here because at present no golden rule for segmentation algorithms exists.
However, a generally well-accepted toolkit for medical image segmentation is ITK. It is available as an open source software package which is widely supported and is thus recommended here [12]. For those not familiar with C++ and installation procedures, a collection of scientific software packages python(x, y) [13] which includes ITK and the visualization toolkit VTK [14] may be easier to use and to install. In addition, ITK and VTK allow reading of the DICOM format and may thus be used for opening the DICOM files of the dataset distribution as well.
For orientation only, a volume reconstruction of the torso from the image stacks of series 301 and 701 was performed using VTK, in order to provide data for the artwork in Figures 3, 4, and 5.
Biosignal data
The biosignals BSP (body surface potential) and MCG (magnetocardiogram) were simultaneously recorded over 100 s with a sampling interval of 1 ms. A respiration signal was recorded simultaneously and is included in the dataset.
The identifiers and positions of the electrode and sensor centers are given in the accompanying spread sheets in the x, y, and z-coordinates of the unique coordinate system introduced in Figure 5.
It should be noted that the MRI series 601 was acquired in CINE mode. For reference, the instances imaged in Figure 6 are marked by red cursors in the subsequent biosignal plots.
BSPM data
The BSP signals recorded contained approximately 120 heart beats in the time series (Additional file 6). They were condensed into a single representative beat shown in Figure 7 in a so-called butterfly plot for all BSP channels. The averaging procedure comprised the temporal overlay of all beats centered at the R peak trigger, a moderate baseline correction and a median filter averaging. A detailed description of such an approach is given, e.g., by Koch et al. [15].
Admittedly, the quality of the BSP signals does not represent the state-of-the-art. This is due to the aim to record them truly simultaneously with the MCG signals. This in turn required the use of distinctly non-magnetic ECG electrodes for the BSPM, otherwise the extremely sensitive magnetic recordings would have been deteriorated by artifacts. GRASS F-E5GH electrodes (Astro-Med, Inc.) have proven to contain a very low level of magnetic contamination. However, their small area and only moderate skin contact stability lead to non-optimal ECG signal quality. We aimed to achieve ultimate MCG quality (see next section), as these signals may be better suited for the intended purpose of forward and inverse model verification. Thus the BSP signal quality had to be a compromise.
From the BSP signals body surface potential maps (BSPM) have been constructed and are shown in Figure 8 for selected time instances marked by the green cursors in Figure 7.
MCG data
Magnetocardiography (MCG) is not as well known as its electric counterpart, the ECG. However, for the purpose of model verification it seems to be superior. It is well documented that the magnetic field generated by electrophysiological activity of the heart muscle is far less deteriorated than the respective electric potential at the torso surface. Contrary to the scalar character of an electric potential field, the magnetic equivalent is a vector field. In addition, MCG recordings are contactless and thus do not contain artifacts due to skin-electrode impedance fluctuations.
The butterfly plot of the MCG is shown in Figure 9 which displays the signals of only 49 channels of the 304 channel system. These stem from the lowest layer of SQUID sensors which measure the z component of the magnetic induction. Figure 9 demonstrates the excellent signal quality of MCG signals. Finally, in Figure 10, the respective MCG maps are shown that correspond to the simultaneously acquired BSPMs shown in Figure 8. The file provided with the attached file folder contains all signals of the system, thus allowing access to the full vector field information (Additional files 7 and 8).