hMRI - A toolbox for quantitative MRI in neuroscience and clinical research.

Neuroscience and clinical researchers are increasingly interested in quantitative magnetic resonance imaging (qMRI) due to its sensitivity to micro-structural properties of brain tissue such as axon, myelin, iron and water concentration. We introduce the hMRI-toolbox, an open-source, easy-to-use tool available on GitHub, for qMRI data handling and processing, presented together with a tutorial and example dataset. This toolbox allows the estimation of high-quality multi-parameter qMRI maps (longitudinal and effective transverse relaxation rates R1 and R2⋆, proton density PD and magnetisation transfer MT saturation) that can be used for quantitative parameter analysis and accurate delineation of subcortical brain structures. The qMRI maps generated by the toolbox are key input parameters for biophysical models designed to estimate tissue microstructure properties such as the MR g-ratio and to derive standard and novel MRI biomarkers. Thus, the current version of the toolbox is a first step towards in vivo histology using MRI (hMRI) and is being extended further in this direction. Embedded in the Statistical Parametric Mapping (SPM) framework, it benefits from the extensive range of established SPM tools for high-accuracy spatial registration and statistical inferences and can be readily combined with existing SPM toolboxes for estimating diffusion MRI parameter maps. From a user's perspective, the hMRI-toolbox is an efficient, robust and simple framework for investigating qMRI data in neuroscience and clinical research.

Introduction

Quantitative MRI (qMRI) finds increasing interest in neuroscience and clinical research because it is not only more sensitive, but also more specific, to microstructural properties of brain tissue such as axon, myelin, iron and water concentration than conventional weighted MRI (Cercignani et al., 2018; Assaf and Basser, 2005; Draganski et al., 2011; Lorio et al., 2014, 2016a; Stüber et al., 2014; Callaghan et al., 2015a). In conventional weighted MRI, the image grayscale values have arbitrary units and the value in a given voxel will depend on a large number of factors, such as the sequence type (e.g. the magnetisation-prepared rapid gradient echo, MPRAGE (Mugler and Brookeman, 1990) versus modified driven equilibrium Fourier transform, MDEFT (Deichmann et al., 2004) for T1-weighted anatomical images), sequence parameters (e.g. repetition time, , echo time, , or flip angle), and hardware effects (e.g. transmit and receive profiles and any scaling factors). In addition, the value will depend on multiple physical tissue properties such as the longitudinal and transverse relaxation times, and , or the proton density, (Helms et al., 2009, 2010). qMRI accounts for these varied effects in order to increase the specificity of the estimated metrics and eventually quantify specific physical tissue properties (Cercignani et al., 2018; Lutti et al., 2010; Weiskopf et al., 2013). In qMRI, the estimated physical value has a direct meaning and is quantified in standardised units (e.g. in seconds) (Koenig et al., 1993). This standardised nature further increases the comparability across sites and time points (Deoni et al., 2008; Weiskopf et al., 2013), which may improve the sensitivity of multi-site studies and longitudinal analyses of development, plasticity and disease progression. A biophysical interpretation of physical qMRI parameters (Callaghan et al., 2015a; Stikov et al., 2015) or a combination of qMRI with biophysical modelling (e.g. Henkelman et al. (2001), Assaf and Basser (2005), or Mohammadi et al. (2015)) enables the in vivo characterisation of key microscopic brain tissue parameters, which previously could only be achieved with ex vivo histology. This concept is called in vivo histology using MRI (hMRI, Weiskopf et al. (2015)).

The estimation of quantitative and semi-quantitative metrics commonly includes one or more of the effective transverse relaxation rate (), the longitudinal relaxation rate (), the proton density (), the magnetisation transfer () saturation and a number of diffusion MRI (dMRI) metrics (Draganski et al., 2011). However, the majority of fundamental and clinical neuroscience studies either refrain from acquiring qMRI data or the quantitative approach is limited to dMRI only. One reason for this might be that standardised qMRI imaging protocols (such as the protocol for the Human Connectome Project (Sotiropoulos et al., 2013)) and processing software (see a summary in Soares et al. (2013)) are readily available for dMRI but less so for other qMRI techniques.

Consequently, the neuroscience and clinical research community lacks a standardised qMRI implementation to handle the wide diversity of data acquisition types and estimate parameters such as , , and saturation, which are sensitive to iron, myelin, and water content in tissue microstructure, and thus provide complementary information to the axonal properties revealed by dMRI. For example, is often estimated using gradient recalled echo data acquired with multiple echo times () to create high resolution maps that show strong contrast between different types of brain tissue (Bernstein et al., 2004). Similarly, and can be derived from two acquisitions varying the excitation flip angle (Wang et al., 1987; Deoni et al., 2005; Deoni, 2007; Schabel and Morrell, 2008; Helms et al., 2008a, 2011; Liberman et al., 2013; Heule et al., 2015; Baudrexel et al., 2016). For accurate estimation of the latter qMRI parameters, an implementation must adequately correct for instrumental biases such as inhomogeneous transmit (Lutti et al., 2010) and receive fields (Volz et al., 2012; Mezer et al., 2016; Lorio et al., 2018). Such correction should be based on the additional mapping of these fields, but the implementation should also provide solutions when these fields have not been measured. For example in clinical settings, MR sequences for estimating instrumental biases are often unavailable due to time, hardware or software constraints. For such studies, the facility to correct for instrumental biases retrospectively using image processing methods that do not rely on additional MRI acquisitions is highly desirable.

A number of open-source tools have been developed to support qMRI use and make qMRI more broadly accessible to neuroscience and clinical research (see for example qMRLab (Cabana et al. (2015), https://github.com/qMRLab/qMRLab), QUIT (Wood (2018), https://github.com/spinicist/QUIT), mrQ (Mezer et al. (2016), https://github.com/mezera/mrQ) and QMAP (https://www.medphysics.wisc.edu/∼samsonov/qmap/). These tools include, to various extents, data acquisition guidelines, tools for protocol simulation and optimisation, a multitude of models and methods for data fitting and estimation of qMRI parameters, and visualisation tools.

While most of these tools focus on model fitting and generation of qMRI maps, the question of spatial and statistical processing of these qMRI maps is not directly addressed. For spatial processing and group level statistical analysis, the established tools are primarily designed for diffusion MRI (see for example TBSS in FSL (Smith et al., 2006) and TRACULA (Yendiki et al., 2011) in FreeSurfer). In addition, a number of custom made tools have been developed based on established neuroimaging software, e.g.: applications of the FreeSurfer surface projection software (Dale and Sereno, 1993; Fischl et al., 1999; Fischl and Dale, 2000) to compare quantitative relaxation and susceptibility data on the cortex (Marques et al., 2017), or usage of the VBM framework (Ashburner and Friston, 2000) to process qMRI data across the whole brain on a voxel-by-voxel basis (Büchel et al. (2004) and Mohammadi et al. (2012) for dMRI metrics, and Table 1 for , , and saturation). One challenge common to all these tools is to find a proper method to locally preserve the qMRI parameters after (non-linear) spatial registration. The methods using the statistical parametric mapping (SPM) framework typically reduce residual misalignment between images by isotropic spatial smoothing. However, applying this framework directly to qMRI data would introduce partial volume effects at tissue boundaries and corrupt the quantitative values. Hence a modified smoothing approach, which aims to achieve within class smoothing only, is preferred. The majority of the studies in Table 1 took advantage of voxel-based quantification (VBQ), an approach introduced by Draganski et al. (2011) and developed for the comprehensive multi-parameter mapping (MPM) approach (Helms et al., 2008b, 2009; Weiskopf et al., 2011, 2013) to correct for potential error introduced by spatial smoothing.

Table 1

Review of studies related to the MPM acquisition protocol using predecessors of the hMRI-toolbox to make inference on myelin (My), iron (Fe), or the volume (Vol) measured by voxel-based morphometry (VBM). Note that all studies used Siemens MRI scanners. Abbreviations: longitudinal relaxation rate, (effective) transverse relaxation rate, (effective) proton density, magnetisation transfer saturation, C controls, P patients, dMRI diffusion MRI, fMRI functional MRI, MEG magnetoencephalography, EEG electroencephalography.

Reference	Predominant tissue feature			qMRI	Subjects	Remarks
	My	Fe	Vol	parameters
Studies demonstrating improved volumetry
Helms et al. (2009)	_	_	✓	MT	49 C	Improved segmentation of deep brain grey matter structures using magnetisation transfer saturation (MT) parameter maps.
Lambert et al. (2013)	_	_	✓	R_1, R2⋆, MT, PD^⋆	34 C	Brainstem segmentation using a modified multivariate mixture of Gaussians based on MPMs.
Studies of myelin and/or iron
Draganski et al. (2011)	✓	✓	✓	R_1, R2⋆, MT	26 C	A fingerprint of age-dependent brain atrophy and underlying microstructural changes in myelin, iron deposits and water.
Lambert et al. (2013)	✓	✓	✓	R_1, R2⋆, MT, PD^⋆	26 C	Characterizing aging in the human brainstem using quantitative multimodal MRI analysis.
Freund et al. (2013)	✓	_	✓	R_1, MT	13 P/18 C	Atrophic and microstructural changes of corticospinal axons and sensorimotor cortical areas observed in the first months after spinal cord injury.
Callaghan et al. (2014)	✓	✓	✓	R_1, R2⋆, MT, PD^⋆	138 C	A whole-brain pattern of age-associated microstructural differences in the asymptomatic population providing insight into the neurobiology of aging.
Lorio et al. (2014)	✓	✓	✓	R_1, R2⋆, MT, PD^⋆	96 C	Deviation of volume changes assessed with VBM using standard T1-weighted or MT maps, attributed to age-related iron changes.
Lorio et al. (2016b)	✓	✓	✓	R_1, R2⋆, MT, PD^⋆	120 C	Impact of microstructural properties of brain tissue-myelination, iron, and water content on automated measures of brain morphology using VBM.
Steiger et al. (2016)	✓	✓	_	R2⋆, MT	31 C	Age-related increase of iron correlated negatively with iron and myelin in the ventral striatum, which predicted individual memory performance.
Whitaker et al. (2016)	✓	_	✓	R_1, MT	297 C	Assessment of layer-specific microstructure in adolescence development and comparison to volume differences.
Carey et al. (2018)	✓	✓	_	R2⋆, MT, R_1, PD^⋆	93 C	Delineation of cortical myelin profiles across cortical depths and mapping of tissue water sensitive parameters (PD^⋆).
Ziegler et al. (2018)	✓	✓	✓	R2⋆, MT	15 P/18 C	Trauma-induced neuroplastic processes in brain and spinal cord including neurodegenerative processes associated with iron and myelin changes.
Studies linking structure and function in the cortex
Dick et al. (2012)	✓	_	_	R_1 (& fMRI)	9 C	Combining structural R_1 mapping with functional tonotopy to reveal their interrelation within the primary auditory areas (A1 and R).
Sereno et al. (2013)	✓	_	_	R_1 (& fMRI)	6 C	Combined structural R_1 mapping and functional retinotopy revealed boundaries visible in the R_1 maps that correspond to recognised retinotopic borders.
Lutti et al. (2014)	✓	_	_	R_1 (& fMRI)	–	Review about using high-resolution mapping of R_1 as an index of cortical myelination.
Helbling et al. (2015)	✓	_	_	R_1 (& MEG)	5 C	Combining non-invasive electrophysiology with MPMs demonstrating microstructure-function relationship.
Carey et al. (2017)	✓	_	_	R_1 (& fMRI)	10 C	Functional and R_1 mapping demonstrate overlap between articulator and R1 myelin proxy maps in pre and post-central regions.
Dick et al. (2017)	✓	_	_	R_1 (& fMRI)	8 C	Functional and R_1 mapping demonstrate strong concordance in the degree of cortical myelination and the strength of tonotopic activation across several auditory cortical regions.
Studies combining MPMs and biophysical models
Callaghan et al. (2015a)	✓	✓	_	R2⋆, MT, R_1	138 C	Combining MPM and a general linear relaxometry model to describe the biophysical relation of R_1 on iron, myelin, and water content.
Mohammadi et al. (2015)	_	✓	_	MT (& dMRI)	38 C	Using MPM-based MT and diffusion-MRI-based tract fibre density as proxies for estimating myelin and axon densities as well as the g-ratio.
Weiskopf et al. (2015)	✓	✓	✓	R2⋆, MT, R_1, PD^⋆	–	Review about in vivo histology using MRI.
Callaghan et al. (2016b)	✓	✓	✓	R2⋆, MT, R_1, PD^⋆	12 C/30 C	Using a biophysical relation between R_1, MT, and R2⋆ to generate synthetic quantitative maps.
Ellerbrock and Mohammadi (2018)	✓	✓	_	MT, PD (& dMRI)	22 C	Comparing different proxies for myelin and fibre volume fractions using MPMs and diffusion MRI.
Edwards et al. (2018)	✓	✓	_	R_1, R2(⋆), MT, PD (& dMRI)	22 C	Review on cortical models for in vivo histology.

A particular instantiation of qMRI developed at 3T, the MPM approach, spans data acquisition, modelling and bias correction of three multi-echo spoiled gradient echo volumes to generate , , , as well as semi-quantitative saturation maps. This framework enables time-efficient whole brain mapping with high isotropic resolution of 800 μm in 27 min (Callaghan et al., 2015b) or reduced MPM protocol (no MT saturation) at 1 mm isotropic resolution in 14 min at 3T (Papp et al., 2016) and has even enabled the acquisition of ultra-high-resolution quantitative maps with 400 μm resolution at 7T (Trampel et al., 2017). This framework has been used in a variety of fundamental and clinical neuroscience studies (Table 1) focussing on: (a) improving the segmentation of deep grey matter structures, (b) evaluating the myelin and iron concentration in the brain and spinal cord, (c) linking structure and function in the cortex, and (d) using the MPM parameter maps as proxies for biophysical tissue models.

In this paper, we present the hMRI-toolbox, a comprehensive open-source toolbox that streamlines all the processing steps required to generate , , and saturation maps and provides appropriate spatial processing for group analyses. The flexible nature of the toolbox makes it applicable to a wide range of data types, from the full MPM protocol to subsets of it, including single contrast echo trains for mapping or variable flip angle data for mapping of and using multi-echo or even single-echo data. The toolbox is embedded in the SPM framework (http://www.fil.ion.ucl.ac.uk/spm/software/spm12/), profiting from the highly accurate spatial registration into a common space and the variety of established statistical inference schemes. The spatial processing part of the toolbox can be applied to any set of rotationally-invariant qMRI maps, including a number of diffusion MRI parameters and all common qMRI metrics.

Background

The MPM protocol

The MPM multi-echo protocol was introduced in Weiskopf and Helms (2008) and Weiskopf et al. (2013) for estimating the longitudinal relaxation rate , the effective transverse relaxation rate , the proton density and the magnetisation transfer and generalises a number of acquisition protocols. It typically involves acquiring six to eight images at different echo times () for each of the PD-, T1-and MT-weighted acquisitions in an RF and gradient spoiled gradient echo sequence (referred to as T1w, PDw and MTw echoes, respectively). The hMRI-toolbox can flexibly deal with a large range of site- and study-specific acquisition schemes, from the full MPM protocol to subsets of it, including single contrast echo trains for mapping or variable flip angle data for mapping of and using multi-echo or single-echo data, comparable to e.g. DESPOT1 (Deoni et al., 2005). Example MPM acquisition protocols can be found at http://hmri.info.

Overview theory of MPM signal model

We give here, and in Fig. 1, a short overview of the theory underlying the qMRI map creation process. For a more detailed outline of the theory and the estimation procedures applied in the hMRI-toolbox see Appendix A.

Fig. 1

Overview of qMRI map generation from the weighted imaging and reference MPM data. The signal S is modelled by the Ernst equation with an exponential decay depending on the echo time . The longitudinal relaxation rate , the effective transverse relaxation rate , the proton density and the magnetisation transfer () saturation are estimated from the data, using approximations for small repetition time and small flip angles α. The transmit and receive bias fields and are used to correct for instrumental biases.

The signal from the multi-echo PDw, T1w and MTw acquisitions can be described by the Ernst equation (Ernst and Anderson, 1966; Helms et al., 2008a, b). The effective transverse relaxation rate can then be derived from the dependence of the signal. The unified description of the multi-echo data from all three contrasts into a single model, denoted as ESTATICS (Weiskopf et al., 2014), provides a more robust estimation of with a higher signal-to-noise ratio compared to separate estimations (Fig. 1a). Using approximations of the signal equations for small repetition time and small flip angles α, the longitudinal relaxation rate , the apparent signal amplitude map (proportional to the proton density ) and the magnetisation transfer can be estimated. At this point (Fig. 1b), the generated maps are biased by transmit (Fig. 1c) and receive (Fig. 1d) field inhomogeneities. The hMRI-toolbox provides correction methods for these bias fields based on specific transmit and receive field measurements or image processing methods. While influences the local flip angle and hence all three (, , ) maps are affected, the RF sensitivity bias field only influences the map (in the absence of subject motion).

The toolbox can also handle the situation where only a subset of data is available. For example, , and can still be estimated when no MTw acquisitions are acquired, alone can be estimated when neither MTw nor T1w acquisitions are available (single multi-echo PDw data). , and saturation maps can be generated from single echo PDw, T1w and MTw images, not requiring multi-echo acquisitions. The theory and map creation tools also encompass the creation of maps from other variable flip angle approaches, such as DESPOT1 (Deoni et al., 2005) or maps from multi-echo data, such as certain susceptibility mapping/weighted imaging approaches.

Methods

Toolbox documentation and installation

The latest version of the toolbox can be downloaded from the hMRI-toolbox page (http://hmri.info) as a zip file (containing the last official release) or by cloning the git repository (https://github.com/hMRI-group/hMRI-toolbox) to keep up-to-date with the latest incremental developments.

Updated documentation is available as a Wiki (https://github.com/hMRI-group/hMRI-toolbox/wiki). It includes installation instructions, an example dataset, a tutorial and a detailed description of the implemented functionalities. Information on releases and versioning, development and contribution guidelines are also provided.

The toolbox has been developed and tested with MATLAB versions 8.0 (R2012b) to 9.3 (R2017b) and SPM12 from version r6906 onwards. Since the hMRI developments will be synchronised with SPM developments, it is recommended to use the most recent SPM release (http://www.fil.ion.ucl.ac.uk/spm/software/spm12/) to benefit from the latest developments.

The hMRI-toolbox is free but copyright software, distributed under the terms of the GNU General Public License as published by the Free Software Foundation (as given in file Copyright.txt). Further details on “copyleft” can be found at http://www.gnu.org/copyleft/.

In particular, the hMRI-toolbox is supplied as is. No formal support or maintenance is provided or implied. Since the toolbox was developed for academic research, it comes with no warranty and is not intended for clinical use.

MPM example dataset

An MPM example dataset from a healthy subject for demonstrating the hMRI-toolbox features was acquired on a 3T Prisma system (Siemens Healthcare, Erlangen, Germany) at the Wellcome Centre for Human Neuroimaging, London, UK. This dataset can be downloaded from http://hmri.info and is described in detail in Callaghan et al. (submitted).

Organisation of the toolbox

The hMRI-toolbox is organised into five main modules (Fig. 2): Configure toolbox, DICOM import, Auto-reorient, Create hMRI maps and Process hMRI maps. While the Configure Toolbox and Process hMRI maps modules can be run for a group of subjects, the DICOM Import, Auto-reorient and Create hMRI maps modules must be run for each subject and each session separately (if several datasets acquired per subject, e.g. in a longitudinal study). A brief description of each module is provided below and details can be found in the Appendices.

Fig. 2

Left: After installation (Section 3.1), the hMRI toolbox can be started from the SPM menu of the batch editor. Five options include toolbox configuration, DICOM import, re-orientation of the data to MNI space, creation and processing of the hMRI maps. Right: Input mask for the map creation part of the toolbox.

Configure toolbox module

The hMRI-toolbox provides the user with a set of default acquisition and processing parameters for most common acquisition protocols without any further customisation. However, customisation is possible and necessary to broaden the toolbox usability to a wider range of protocols, scanners and vendors. Therefore, the Configure toolbox module allows the user to select either standard or customised default parameters, to match the user's own site- or protocol-specific setup and to be used in the subsequent modules (see details in Appendix C.1).

DICOM import module

DICOM import is a tool to convert DICOM data into NIfTI files. During conversion, the whole DICOM header is stored as JSON-encoded metadata in a file along side the NIfTI images. This feature is implemented in SPM12 from release r7219 (November 2017), following the hMRI-toolbox implementation. Note that Philips-specific rescaling factor (Chenevert et al., 2014) is applied at conversion, starting from version v0.2.0 of the hMRI-toolbox and release r7487 of SPM12. The hMRI-toolbox also provides metadata handling functionalities to retrieve parameter values and store processing parameters in the JSON-encoded metadata file (see Table 3 for an example of processing information stored as metadata). Detailed description of the DICOM import module, the metadata handling tools and BIDS compliance aspects (Gorgolewski et al., 2016) is provided in Appendix B.

Auto-reorient module

The reorientation of the images towards a standard pose, setting the anterior commissure at the origin and both anterior and posterior commissure (AC/PC) in the same axial plane, as defined in MNI space (Mazziotta et al., 1995, 2001a,b), is a common step that increases the consistency in individual head positions prior to normalisation or segmentation. For example, SPM's segmentation (Ashburner and Friston, 2005) is sensitive to the initial orientation of the images. Therefore, Auto-reorient provides a simple tool for automatically and uniformly reorienting a set of images prior to any further processing including multi-parameter map calculation. Note, that the reorientation modifies the orientation information in each image header, but the reoriented images are not resliced. For more details, see Appendix C.2.

Create hMRI maps module

The Create hMRI maps module computes quantitative as well as semi-quantitative estimates of , , and saturation from unprocessed multi-echo T1w, PDw and MTw spoiled gradient echo acquisitions. The map creation module corrects the qMRI estimates for spatial receive (Section 3.7.3) and transmit (Section 3.7.2) field inhomogeneities.

Multi-parameter input images

The module takes the (possibly reoriented) series of multi-echo spoiled gradient echo images as input for the creation of quantitative as well as semi-quantitative maps (Fig. 3) of , , and saturation (Helms et al., 2008a, b; Weiskopf et al., 2013, 2014) as described in the Background Section 2 and Appendix A. The number and quality of the output maps depends on the contrasts (PDw, T1w, MTw) and number of echoes available, and on the availability of additional bias field measurements. A single multi-echo PDw contrast allows for the calculation of a single map. If T1w images are additionally provided, and maps will also be generated. The map estimation requires the acquisition of additional MTw images. The map creation also involves optional correction for ( transmit) bias field (Section 3.7.2) and the (RF receive sensitivity) bias field (Section 3.7.3) as well as spoiling imperfections (Yarnykh, 2010; Preibisch and Deichmann, 2009).

Fig. 3

Map creation workflow illustrated for the MPM example dataset (Callaghan et al., submitted). The MPM protocol includes three multi-echo spoiled gradient echo scans with predominant T1-, PD- and MT-weighting achieved by an appropriate choice of the repetition time, the flip angle and the off-resonance MT pulse. Optional RF transmit () and receive () field measurements can be added to the protocol, improving the quality of instrumental bias correction in the MPM maps. Alternatively, these reference measurements can be, to a limited extent, replaced by dedicated image processing steps that are provided by the toolbox. The map creation module generates , , and maps. For each map, a JSON metadata file is created, which contains information about the processing pipeline of each image (see example in Table 3).

B1 (transmit) bias correction

The map creation module includes the determination of transmit bias field maps ( expressed in p.u. Of the nominal flip angle) for transmit bias correction of the quantitative data. Several methods are implemented. Depending on the choice of the specific method the GUI requires the user to provide adequate input files. Further details on the supported correction methods can be found in Appendix C.3 and the respective original publications (Lutti et al., 2010, 2012; Weiskopf et al., 2013, 2011; Yarnykh, 2007; Chung et al., 2010).

Receiver RF sensitivity bias correction

Three options are available to correct for RF receive sensitivity bias () within the Create hMRI maps module. Two of them rely on measured RF sensitivity maps (Single or Per contrast options) while the third method is data driven (Unified Segmentation option: no input sensitivity map required). Although not recommended, it is possible to disable RF sensitivity correction altogether by selecting the None option. While options Single and Unified Segmentation assume that the sensitivity profile is consistent between contrasts (i.e. small inter-contrast subject movement is assumed), the Per contrast option accounts for inter-contrast variation in RF sensitivity profile due to larger subject motion (Papp et al., 2016). Details on the different RF sensitivity bias correction methods can be found in Appendix C.4 and in the respective publications (Papp et al., 2016; Weiskopf et al., 2013).

Output

By default, the estimated quantitative maps are output into a Results subdirectory within the folder of the first PDw echo. Alternatively, a user-defined folder for the output of the toolbox can be selected in which the Results directory will be created. The estimated qMRI maps are saved in the Results directory, with supplementary files are output in the Results/Supplementary subfolder. The basename for all qMRI maps is derived from the first echo of the PDw image series, see Table 2 for brief description. If data is reprocessed, a new sub-folder is created.

Table 2

Output files from the Create hMRI maps module using the SE/STE mapping and per-contrast RF sensitivity bias correction.

Results directory	Description
<firstPDfileName>_MTsat.[nii|json]	Estimated MT saturation map in p.u.
<firstPDfileName>_PD.[nii|json]	Estimated PD map in p.u.
<firstPDfileName>_R1.[nii|json]	Estimated R_1 map in s^−1
<firstPDfileName>_R2s_OLS.[nii|json]	Estimated R2⋆ map in s^−1 (ESTATICS) with ordinary least squares fit
Results/Supplementary directory	Description
hMRI_map_creation_rfsens_params.json	RF sensitivity bias correction parameters (measured sensitivity maps)
hMRI_map_creation_b1map_params.json	B_1 transmit map estimation: acquisition and processing parameters
hMRI_map_creation_job_create_maps.json	Create hMRI maps: acquisition and processing parameters
hMRI_map_creation_mpm_params.json	Acquisition and processing parameters used for the current job
hMRI_map_creation_quality_assessment.json	Quality assessment results (Appendix D)
<firstSESTEfileName>_B1map.[nii|json]	Estimated B_1 bias field f_T map (p.u.)
<firstSESTEfileName>_B1ref.[nii|json]	Anatomical reference for B_1 bias field correction
<firstPDfileName>_MTw_OLSfit_TEzero.[nii|json]	MTw echoes extrapolated to TE=0
<firstPDfileName>_PDw_OLSfit_TEzero.[nii|json]	PDw echoes extrapolated to TE=0
<firstPDfileName>_R2s.[nii|json]	Estimated R2⋆ map from simple exponential fit (PDw echoes)
<firstPDfileName>_T1w_OLSfit_TEzero.[nii|json]	T1w echoes extrapolated to TE=0

Process hMRI maps module

The Process hMRI maps module provides dedicated tools and tissue probability maps for the spatial processing of quantitative MRI maps based on the corresponding SPM framework. The spatial processing pipeline for hMRI data relies on three main operational steps (Fig. 4): (1) segmentation (Ashburner and Friston, 2005; Draganski et al., 2011; Lorio et al., 2016a), (2) non-linear spatial registration into common space (Ashburner (2007)) and (3) tissue-weighted smoothing (Draganski et al. (2011)), using three different sub-modules that are further detailed in the Appendix, Table 4. Furthermore, a fully integrated processing pipeline is provided as an additional sub-module to facilitate standard data processing without the need to combine the individual steps in this module. Details on the three sub-modules and the integrated pipeline are provided in Appendix C.5.

Fig. 4

Overview of the spatial processing module. It consists of three steps: (1) segmentation, (2) highly parametrised non-linear spatial registration, and (3) tissue-weighted smoothing. The segmentation step (1) uses novel tissue-probability maps (TPMs), designed to take advantage of the better contrast of the MPMs for improved segmentation (exemplified for the deep-grey matter by the arrow in (1)). The non-linear spatial registration step into common space (2) reduces inter-individual anatomical differences (exemplified for maps of subject one and two, S1 and S2, respectively). To further reduce residual anatomical differences (see magnification boxes in (2)) and enhance statistical inference, the qMRI maps can be spatially smoothed (3) using the voxel-based quantification (VBQ) smoothing procedure. As compared to Gaussian smoothing, VBQ smoothing avoids bias in the qMRI maps (e.g. see arrows in (3), highlighting a rapid decline in values at tissue boundaries only after Gaussian smoothing. The VBQ smoothing is detailed in Eq. (C.3) and Draganski et al. (2011)). The sub-figure (1) has been adapted from (Lorio et al., 2016a), sub-figures (2) and (3) from Mohammadi and Callaghan (2018).

Statistical analysis

The standard SPM statistical analysis and modelling approaches such as mass-univariate General Linear Modelling can be applied to the spatially processed maps, see, e.g., Draganski et al. (2011) and Freund et al. (2013). Additionally, the multiple parameter maps lend themselves to multi-variate analyses approaches as well (Draganski et al., 2011).

Discussion and outlook

This paper introduced the hMRI-toolbox, which is embedded in the SPM framework and allows for the estimation and processing of four quantitative MRI parameter maps: the longitudinal and effective transverse relaxation rates and , the proton density and the (semi-quantitative) magnetisation transfer saturation . This introduction includes a comprehensive summary of the MPM signal model as well as the currently available correction methods for the various bias sources that, if not corrected for, might impair the quantification. Finally, the processing steps for a dedicated SPM analysis of quantitative MRI maps (denoted as the VBQ approach) were presented, correcting for the potential partial volume effects introduced by spatial smoothing.

The name of the toolbox (h-MRI) originates from the concept of in vivo histology of tissue microstructure using MRI (Weiskopf et al., 2015). Hereby, the quantitative parameter maps generated with this toolbox provide key input parameters for biophysical models that are designed to non-invasively estimate specific microstructural tissue properties (see Table 1 for example studies).

Considerations and interpretation of the MPM approach. While the signal model (see Appendix A for details) used in this toolbox is based on the Ernst equation and thus provides a comprehensive means of calculating a set of physical quantitative (and semi-quantitative) parameters, including , , and , we would like to emphasise that more sophisticated models can be derived to relate the parameters more directly to the underlying biophysical mechanisms and tissue characteristics.

The contrast provided by the metric is associated with different realisation of iron deposits (for a review, see e.g. Edwards et al. (2018)), myelination of axons (Marques et al., 2017), their orientation (Oh et al., 2013) and chemical exchange (Does, 2018). Multi-compartment models can be used, e.g., to separately describe the orientation-dependence of myelinated fibre pathways in parameter maps (e.g., Lee et al. (2016); Alonso-Ortiz et al. (2018); Wharton and Bowtell (2012)).

At a given field strength, the contrast is determined by the micro-structural tissue properties such as the local mobility of water molecules, the macromolecular content and the local concentration of paramagnetic ions such as iron or gadolinium-based contrast agents. It has been shown that depends within limits linearly on these tissue properties (Fatouros and Marmarou, 1999; Fatouros et al., 1991; Kaneoke et al., 1987; Shuter et al., 1998; Donahue et al., 1994; Kamman et al., 1988; Fullerton et al., 1982; Gelman et al., 2001). The and metrics can be used to describe the dependence of on these components (Callaghan et al., 2015a).

saturation is a proxy measure of the bound-pool water fraction (Helms et al., 2008b). It provides information about the macromolecular content of the micro-structural environment and is often used as a marker for myelin content, see e.g. Freund et al. (2013) and Callaghan et al. (2014). Under equivalent conditions for the off-resonance pre-pulse the same saturation values are expected. However, this will not be the case if the properties of the pre-pulse are changed across measurements. Therefore, we refer to the saturation measure as being semi-quantitative. The saturation map differs from the commonly used ratio (MTR; percent reduction in steady state signal) by explicitly removing the bias introduced by the spatially varying relaxation time and -transmit field (Helms et al., 2008b). Additional minor corrections for transmit field inhomogeneity in the maps were applied as described in Weiskopf et al. (2013). The reduced spatially varying bias leads, e.g., to a higher contrast in deep brain structures than MTR and to reduced variance in the data (Callaghan et al., 2016b). Note that the saturation measure does not only depend on the bound-pool fraction but also on the exchange between the bound and free water pools (see e.g. Battiston and Cercignani (2018)). A more direct measure of the bound-pool fraction is provided by quantitative (), which requires more time-consuming data acquisition, typically limiting the map's spatial resolution (e.g., 2 mm isotropic in Stikov et al. (2011)).

Sources of bias in qMRI and limitations to bias corrections. Correction of instrumental characteristics and artefacts is an essential prerequisite for quantitative MRI. Sources of biases and artefacts include primarily transmit and receive fields, imperfect spoiling, -bias and head motion. The artefact correction methods provided in the hMRI-toolbox are highly flexible, offering solutions to process reference measurements (e.g. transmit/receive field measurements carried out with a number of customised or product sequences) to correct for instrumental artefacts, as well as achieve optimal results even when no adequate measurements are available.

The ideal imaging protocol includes dedicated measurements of transmit and receive field inhomogeneities, which are typically based on customised sequences (Fig. 5a–d for RF receive bias correction, and Fig. 6a for transmit bias correction). When customised sequences are not available, a standard spoiled gradient echo product sequence can be used to acquire data for low-resolution receive field mapping. Similarly, when not available, customised reference and calibration sequences such as 3D_EPI (Lutti et al., 2010, 2012) or 3D_AFI (Yarnykh, 2007) for correction of transmit effects can be replaced by manufacturer's service sequences such as TFL_B1_map (Chung et al., 2010) or RF_map (as examples on Siemens scanners). The toolbox provides the option to process data from several different transmit bias field mapping techniques or to use transmit bias field maps pre-computed outside the toolbox (see section 3.7.2 and Appendix C.3).

Fig. 5

Comparison of different available RF sensitivity bias field correction methods (Appendix C.4) demonstrated on the MPM example dataset (Section 3.2) and difference maps. (a) map calculated using the reference correction method (), following the approach in Papp et al. (2016), requiring three sensitivity maps, each acquired directly before the respective PDw, MTw and T1w contrasts. (b–d) Difference between and maps calculated from PDw and T1w images corrected by a single sensitivity map, acquired directly before the PDw images (b: ), the MTw images (c: ) and the T1w images (d: ), respectively. Due to the large overt movement preceding the acquisition of the MTw images and corresponding RF sensitivity measurement (see Callaghan et al. (submitted)), there is a large discrepancy between head position for that specific RF sensitivity measurement on the one hand and head positions for the PDw and T1w images used to generate the map on the other hand. As a result, errors in (c) are much larger than in (b) and (d). (e) Difference between and the map corrected for RF sensitivity bias using the Unified Segmentation approach (). The body coil sensitivity profile, not corrected for in the reference method, modulates the difference map in (e). (f) Sagittal view of map in (a) depicting the position of the slice shown in (a–e).

Fig. 6

Comparison of two different available transmit field correction methods demonstrated on longitudinal relaxation rate () maps. (A) The map is depicted after using the reference correction method, following the mapping approach by Lutti et al. (2012). (B) The difference between the reference map and the map derived with UNICORT transmit bias correction (image processing method, no additional measurement required, see Weiskopf et al. (2011) for details). (C) The difference between the reference map and the map derived without transmit bias correction. (B) Shows a comparatively small residual modulation across the slice while (C) is strongly biased by the transmit inhomogeneity across the slice. The slice location in (A–C) is depicted in Fig. 5f.

The measured transmit and receive fields can be affected by diverse sources of error leading to imperfect corrections. For example, residual misalignment between measured receive (or transmit) field and the spoiled gradient echo images can be one reason for such imperfections. In particular, when between-contrast (PDw, T1w, MTw) motion is large, discrepancies between head position for a single receive (or, to a lesser extent, transmit) field measurement and head position for all or some of the spoiled gradient echo images lead to additional motion-related bias in the quantitative maps (Fig. 5). In such a case, a per-contrast RF receive sensitivity measurement is preferable and can account for the between-contrast dynamic variation (see Appendix C.4, Receive field sensitivity measurements and Fig. 5). However, measured RF receive sensitivity maps such as described in Appendix C.4 can also suffer from residual modulations by the receive field of the body coil, which serves as a reference and whose inhomogeneity is not accounted for. Such modulation cannot be directly corrected for using the measured transmit field of the body coil at 3T and higher fields due to the non trivial applicability of the reciprocity principle at such field strengths (Hoult, 2000). As a result, when no large between-contrast motion is observed, RF sensitivity bias correction using the data driven receive field estimation (described in Appendix C.4) may prove more effective altogether. Such a data driven method could also be applied (with specific optimisation of the US regularisation parameters to the body coil's receive field profile) to correct for the above residual body coil receive field modulation. Finally, at 7T where RF body coils are not available, the currently implemented RF sensitivity measurements and bias correction approach are not applicable. All the above aspects are active fields of investigation, optimisation and validation that are also an integral part of the future developments of the toolbox. Similarly, transmit field mapping techniques can be inaccurate. For a comparison of the frequently used -transmit field mapping techniques and a description of their respective accuracy and sources of uncertainty we refer to Lutti et al. (2010) and Pohmann and Scheffler (2013).

When no transmit and/or receive field inhomogeneity maps have been measured, which often happens in clinical settings due to time constraints, the toolbox provides the option to use image processing methods based on the Unified Segmentation approach (Ashburner and Friston, 2005) for transmit bias correction (UNICORT, Weiskopf et al., 2011; see Fig. 6b) or RF sensitivity bias correction (Fig. 5e). The Unified Segmentation approach takes advantage of the fact that bias corrections can be applied post hoc in good approximation for small read-out flip angles and short (Helms et al., 2008a). This requires no additional acquisition time but produces quantitative maps of lesser accuracy with some residual receive and/or transmit field modulation (Weiskopf et al., 2011; Baudrexel et al., 2016) compared to a correction with measured references. Note that the Unified Segmentation approach, whether applied for transmit (UNICORT) or RF sensitivity bias correction, has been optimised for the Siemens TIM-TRIO MR system using the body RF coil for transmission and the 32-channel receive head coil (Weiskopf et al., 2011). The corrections will perform appropriately for coils with similar transmit or receive field profiles, but might require further adjustments otherwise (see Weiskopf et al. (2011) for UNICORT optimisation). For more details, see (Callaghan et al., 2016a) and the hMRI wiki (the latter also provides further information on customized usage).

The proposed MPM protocol uses RF and gradient spoiling to minimise undesired transverse net magnetisation. Imperfect spoiling, which depends on the precise sequence protocol settings, can leave a residual bias in the map if no further correction is used (Preibisch and Deichmann, 2009; Yarnykh, 2010). For specific MPM protocols using the customised sequences, the hMRI-toolbox provides a correction for imperfect spoiling, see Eq. (A.15) in Appendix A.5. By default this correction is disabled but can be enabled through the toolbox customisation provided by the Configure toolbox module (Appendix C.1 and the hMRI-toolbox Wiki).

The estimation of maps can be biased by relaxation effects if not accounted for (Eq. (A.4) in Appendix A). Two correction methods, based on extrapolation of the data to (Ellerbrock and Mohammadi, 2018) or relying on the estimated maps respectively, are implemented in the toolbox (Balteau et al., 2018). These correction methods require a number of echoes to be acquired for a robust fit of the exponential decay to derive magnitude images and maps. In the case of a single-echo variable flip angle dataset, the -weighting correction cannot be applied and the estimated and (to a much lesser extent) the and maps will be biased by -modulations. Also, maps are biased in areas with severe susceptibility artifacts (Yablonskiy, 1998). Note that the JSON metadata file associated with the respective PD parameter map contains information about the processing steps and thus of potential modulation.

Head motion is widely recognised as a major source of artefacts in MR images, with severe consequences for quantitative MRI and morphological measures of the brain (Callaghan et al., 2015b; Weiskopf et al., 2014; Reuter et al., 2015). While quantitative measures of image quality have been introduced, visual inspection remains the most common means of rating data quality despite its limited sensitivity and inter-rater variability (Rosen et al., 2018). The hMRI-toolbox provides summary measures of head motion within and between the acquisitions of each image volume (intra- and inter-scan motion) (Castella et al. (2018) and Appendix D). The provided index of intra-scan motion has been tested against the history of head motion, recorded in real-time during the scans (Castella et al., 2018). Note that these intra- and inter-scan head motion measures could potentially be combined to guide toolbox users to objectively classify their data according to quality, for example to exclude or downweight poor-quality data of individuals in a statistical group analysis.

Spatial processing pitfalls. Since spatial processing in the hMRI-toolbox is embedded in the SPM framework, it is subject to the same limitations as any typical VBM study, including spatial normalisation accuracy, segmentation errors and partial volume effects (Ashburner and Friston, 2000; Ridgway et al., 2008; Focke et al., 2011). Auto-reorient is an option that can help improve the segmentation. However, it has to be done carefully. Poor signal-to-noise ratio, contrast-to-noise ratio, or outliers in the MPM input images may impair the reorientation procedure. In general, it is good practice to visually inspect the results of the hMRI pipeline to detect any obviously suspicious results. Piloting the processing pipeline using a batch for a single healthy subject dataset, combined with a careful check of the log files (in Results/Supplementary, metadata including summary description as shown in Table 3), and comparison with the computed MPM maps from the sample dataset, is recommended.

Residual misalignments between qMRI maps of individual participants in common space will be present despite the high degree of spatial correspondence that can be achieved by the non-linear warping algorithms available in SPM, e.g., by DARTEL (Ashburner, 2007). These misalignments are typically reduced in the VBM-framework by spatial smoothing. To correct for the partial volume effects at tissue boundaries that can be introduced by spatial smoothing, the hMRI-toolbox provides the dedicated smoothing approach that has been described in Fig. 4.

The choice of the appropriate smoothing kernel and its performance compared to alternative methods (e.g. TBSS (Smith et al., 2006) or TSPOON (Lee et al., 2009)) is still a subject of active research. An alternative method for reducing spatial misregistration in the cortex might be surface-based registration algorithms (Davatzikos and Bryan, 1996; Drury et al., 1996; Thompson and Toga, 1996; Fischl et al., 1999). However, a comparison study by Klein et al. (2010) between volume-based and surface-based registration methods could not demonstrate clear superiority of one or the other approach. In fact this is an active area of research to better understand the relative benefits and pitfalls of each approach (Canna et al., 2018).

hMRI-toolbox for different MRI scanner platforms. The first task for interested users of the hMRI-toolbox is the setup of the MPM acquisition protocol. To facilitate standardisation, a set of example protocols for the customised MPM sequences on Siemens platforms as well as standard sequences available on Siemens and Philips platforms are provided on the http://hmri.info website. Those protocols take advantage from the fact that MPM sequences primarily rely on multi-echo spoiled gradient echo sequences that are available on all modern MRI scanners. Even though the experience with implementing the specific MPM protocol on the Philips platforms is limited, first important steps have been achieved (Lee et al., 2017, 2018), and the MPM framework together with the hMRI-toolbox will be used in a multi-site clinical trial (NISCI trial, (Seif et al., 2018),) including Philips and Siemens MR systems.

Multi-scanner and multi-vendor data sets will require adjustments in terms of data handling and processing with the hMRI-toolbox. Most MPM studies up to now were carried out on Siemens MRI scanners using customised MPM sequences. Consequently, the toolbox is optimised for this scenario. New issues might arise when implementing the MPM protocol and using data from other vendors or conventional product sequences. For instance, different MT pre-pulse implementations will lead to changes of the saturation map, which will require appropriate inter-scanner calibration (Volz et al., 2010; Seif et al., 2018). Moreover, not every DICOM to NIfTI conversion software appropriately handles image intensity scaling, as reported e.g. for Philips data (Chenevert et al., 2014), leading to spurious intensity differences affecting the quantification. Making the hMRI-toolbox data formats fully BIDS compliant (Gorgolewski et al., 2016) by defining the ontology of acquisition parameters necessary for the creation of the quantitative maps (see Appendix B.5) further supports the use on multiple platforms and vendors, and is a high priority of the ongoing developments.

Applicability of the toolbox to different MRI platforms also involves ultra-high field MR systems. With the fast progress of ultra-high field scanners (7T and higher) on the MRI market, high-resolution data will become more routinely available and provide access to e.g. laminar-specific information (see 400 μm new generation MPM, Trampel et al. (2017)), while also posing new challenges (e.g. field inhomogeneities) that the hMRI-toolbox will have to address.

Future directions. The hMRI-toolbox has been developed as a scientific collaborative project. As such its developments aim at making it broadly available, capitalising on its flexible and open-source implementation, and adjusting to data sets acquired on multiple MRI platforms (see hMRI-toolbox for different MRI scanner platforms above).

Sensitivity to small inter-individual changes of microstructure (e.g. plasticity) and the variation of change across subjects (e.g. in development, see https://www.biorxiv.org/content/early/2018/07/26/328146) is another challenge for longitudinal qMRI. To that end, the bias of the qMRI estimates in the presence of motion has been investigated (Weiskopf et al., 2014; Callaghan et al., 2015a; Castella et al., 2018). Retrospective robust estimation of parameters (outlier rejection) has been suggested in Weiskopf et al. (2014) and could be implemented in future releases of the hMRI toolbox. Robustness must also be considered in parallel to the spatial resolution versus sufficient SNR level tradeoff to improve the sensitivity of the technique to small developmental changes and plasticity. Thus, spatially adaptive noise removal methods (Tabelow et al., 2016) along with appropriate handling of the Rician bias problem (Polzehl and Tabelow, 2016; Tabelow et al., 2017) are important future improvements of the hMRI-toolbox.

Currently, quality assessment (QA) is provided in the hMRI-toolbox via a set of indicative parameters that should be used at the user's discretion (Appendix D). Future work will focus on further validating and implementing an automated QA of the raw data and generated maps, providing figures from representative populations and protocols.

Extensions of the hMRI-toolbox can easily fit within its modular implementation. As short term future additions, the following three modules and extensions are considered. An additional module that efficiently calculates the protocol-specific correction parameters required to account for imperfect spoiling (Appendix A.5) is planned. Quantitative susceptibility mapping (QSM), taking advantage of the existing phase images acquired with the MPM protocol (Acosta-Cabronero and Callaghan, 2017; Acosta-Cabronero et al., 2018), is a second extension. Finally, as suggested by the ”h” in hMRI, adding new biophysical models that take advantage of the multi-contrast MRI data and generated qMRI maps for in vivo histology (Weiskopf et al., 2015) is another priority of future developments. An example for such a direct extension of the hMRI-toolbox could be the MR g-ratio model (Stikov et al., 2015; Mohammadi et al., 2015; Ellerbrock and Mohammadi, 2018). The MR g-ratio (the ratio between inner and outer diameter of a myelinated axon) is a geometrical microstructural tissue property that can be derived by combining myelin-sensitive qMRI maps from the hMRI-toolbox (e.g. or maps) with the axonal-sensitive maps obtained with existing SPM tools (e.g. the ACID toolbox (Mohammadi et al., 2012; Tabelow et al., 2015; Ruthotto et al., 2013; Mohammadi et al., 2014) and the DTI & Fiber tools (Reisert et al., 2013)).

Conclusion

The hMRI-toolbox is a highly flexible software package that provides a computationally time-efficient, robust and simple framework to generate and process qMRI parameter maps sensitive to myelin, iron, and water content. It profits from the powerful and easy-to-use spatial and statistical analysis tools in SPM, and can be readily combined with existing SPM tools for quantitative estimation of parameter maps sensitive to complementary information such as axonal properties. The ongoing developments address the use of open-science data formats and extensions into biophysical models for direct microstructure mapping. As such, the hMRI-toolbox is a comprehensive and readily extendable tool for estimating and processing qMRI data for neuroscience and clinical research.

Table 3

JSON Metadata example: Metadata are stored as a Matlab structure in the JSON file associated to each output (NIfTI) image. The metadata structure can easily be retrieved using the following SPM/Matlab command: metadata = spm_jsonread(). The structure includes processing parameters, dependencies of the result on other input images as well as a summary description of each output image. As an example, the image type of each output image used to generate Fig. 6 is provided in the Table. The highlighted lines (bold font) describe the transmit bias correction, illustrating the user-friendly traceability of the processing steps used to generate an image. The same information can be read directly from the JSON file using a text editor.

Output map	metadata.history.output.imtype
(A) R_1,Ref	R1 map [s-1]
	- B1+ bias correction using provided B1 map (i3D_EPI)
	- RF sensitivity bias correction based on a per-contrast sensitivity measurement
(B) R_1,UNICORT	R1 map [s-1]
	- B1+ bias correction using UNICORT
	- RF sensitivity bias correction based on a per-contrast sensitivity measurement
(C) R_1,None	R1 map [s-1]
	- no B1+ bias correction applied
	- RF sensitivity bias correction based on a per-contrast sensitivity measurement

Source image	Recommended template
First T1w echo	SPM/canonical/avg152T1.nii
First PDw echo	SPM/canonical/avg152PD.nii
First MTw echo	SPM/canonical/avg152PD.nii

Table 4

Output files from the Process hMRI maps modules. The file names are based on the file name of the map or image used as input to the Segmentation, Dartel > Run Dartel (create Templates), Dartel > Normalise to MNI space, and Smoothing steps. Note that wc*[nii|json]. images can be obtained either with the simple warp obtained with the Unified Segmentation or following Dartel with the Dartel > Normalise to MNI space module.

File name	Description
Unified Segmentation (US) (Proc. hMRI -> Segmentation)
c1<segmInputFileName>.[nii|json]	Tissue class 1 in subject space. Prefixes c1/c2/c3 correspond to GM/WM/CSF respectively.
rc1<segmInputFileName>.[nii|json]	Tissue class 1 in subject space, resliced and imported for Dartel processing. Dartel usually relies on rc1/rc2 images (GM/WM) only.
wc1<segmInputFileName>.[nii|json]	Tissue class 1 warped into MNI space, with the simple warp obtained with the US. Prefixes wc1/wc2 correspond to GM/WM respectively.
mwc1<segmInputFileName>.[nii|json]	Tissue class 1 warped into MNI space, with the simple warp obtained with the US, and modulated by the determinant of the Jacobian, i.e. accounting for local change of volume. Prefixes mwc1/mwc2 correspond to GM/WM respectively.
y_<segmInputFileName>.nii	Deformation field, i.e. warps, obtained from US
<segmInputFileName>_seg8.mat	Segmentation and warping parameters, obtained from US
Diffeomorphic registration (Proc. hMRI -> Dartel -> Run Dartel (create Templates))
u_rc1*_Template.nii	Flow field image, one per subject, estimated by Dartel from the rc1/rc2 images.
Template_*.nii	7 template images, numbered from 0 to 6, created by Dartel from the rc1/rc2 images of all the subjects.
Diffeomorphic registration (Proc. hMRI -> Dartel -> Normalise to MNI space)
w*<normInputFileName>.[nii|json]	Image warped into MNI space following Dartel, using the estimated flow field and an affine transformation. This would be typically a qMRI map that should not be modulated to account for volume changes.
mw*<normInputFileName>.[nii|json]	Image warped into MNI space and modulated by the determinant of the Jacobian, i.e. accounting for local change of volume. This would typically be any tissue probability map (i.e. image with a measure whose total amount over the brain volume should be preserved) to be used after smoothing for a VBM analysis.
Tissue-weighted smoothing (Proc. hMRI -> Smoothing)
wap1_<smooFileName>.[nii|json]	Tissue-weighted smoothing for tissue class 1.
s*<FileName>.[nii|json]	(Any) image smoothed with a standard Gaussian filter (here for comparison, not in hMRI).