Images and other data

Format handling in MRtrix3

MRtrix3 provides a flexible data input/output back-end in the shared library, which is used across all applications. This means that all applications in MRtrix3 can read or write images in all the supported formats - there is no need to explicitly convert the data to a given format prior to processing.

However, some specialised applications may expect additional information to be present in the input image. The MRtrix .mif/.mih formats are both capable of storing such additional information data in their header, and will hence always be supported for such applications. Most image formats however cannot carry additional information in their header (or at least, not easily) - this is in fact one of the main motivations for the development of the MRtrix image formats. In such cases, it would be necessary to use MRtrix format images. Alternatively, it may be necessary to provide the additional information using command-line arguments (this is the case particularly for the DW gradient table, when providing DWI data in NIfTI format for instance).

Image file formats are recognised by their file extension. One exception to this is DICOM: if the filename corresponds to a folder, it is assumed to contain DICOM data, and the entire folder will be scanned recursively for DICOM images.

It is also important to note that the name given as an argument will not necessarily correspond to an actual file name on disk: in many cases, images may be split over several files. What matters is that the text string provided as the image specifier is sufficient to unambiguously identify the full image.

Coordinate system

All MRtrix applications will consistently use the same coordinate system, which is identical to the NIfTI standard. Note that this frame of reference differs from the DICOM standard (typically the x & y axis are reversed). The convention followed by MRtrix applications is as follows:

dimensional description
0 (x) increasing from left to right
1 (y) increasing from posterior to anterior
2 (z) increasing from inferior to superior

All coordinates or vector components supplied to MRtrix applications should be provided with reference to this coordinate system.

Multi-file numbered image support

It is possible to access a numbered series of images as a single multi-dimensional dataset, using a syntax specific to MRtrix. For example:

$ mrinfo MRI-volume-[].nii.gz

will collate all images that match the pattern MRI-volume-<number>.nii.gz, sort them in ascending numerical order, and access them as a single dataset with dimensionality one larger than that contained in the images. In other words, assuming there are 10 MRI-volume-0.nii.gz to MRI-volume-9.nii.gz, and each volume is a 3D image, the result will be a 4D dataset with 10 volumes.

Note that this isn’t limited to one level of numbering:

$ mrconvert data-[]-[].nii combined.mif

will collate all images that match the data-number-number.nii pattern and generate a single dataset with dimensionality two larger than its constituents.

Finally, it is also possible to explicitly request specific numbers, using a Number sequences and floating-point lists within the square brackets:

$ mrconvert data-[10:20].nii combined.mif

Data types

MRtrix applications can read and write data in any of the common data types. Many MRtrix3 commands also support the -datatype option to specify the data type for the output image. For example:

$ mrconvert DICOM_images/ -datatype float32 output.nii


Not all image formats support all possible datatypes. The MRtrix image file formats are designed to handle all of the possibilities listed below, while other image formats may only support of subset. When a data type is requested that isn’t supported by the image format, a hopefully suitable alternative data type will be used instead.

Below is a list of the supported data types and their specifiers for use on the command-line. Note that MRtrix is not sensitive to the case of the specifier: uint16le will work just as well as UInt16LE.

Specifier Description
Bit bitwise data
Int8 signed 8-bit (char) integer
UInt8 unsigned 8-bit (char) integer
Int16 signed 16-bit (short) integer (native endian-ness)
UInt16 unsigned 16-bit (short) integer (native endian-ness)
Int16LE signed 16-bit (short) integer (little-endian)
UInt16LE unsigned 16-bit (short) integer (little-endian)
Int16BE signed 16-bit (short) integer (big-endian)
UInt16BE unsigned 16-bit (short) integer (big-endian)
Int32 signed 32-bit int (native endian-ness)
UInt32 unsigned 32-bit int (native endian-ness)
Int32LE signed 32-bit int (little-endian)
UInt32LE unsigned 32-bit int (little-endian)
Int32BE signed 32-bit int (big-endian)
UInt32BE unsigned 32-bit int (big-endian)
Float32 32-bit floating-point (native endian-ness)
Float32LE 32-bit floating-point (little-endian)
Float32BE 32-bit floating-point (big-endian)
Float64 64-bit (double) floating-point (native endian-ness)
Float64LE 64-bit (double) floating-point (little-endian)
Float64BE 64-bit (double) floating-point (big-endian)
CFloat32 complex 32-bit floating-point (native endian-ness)
CFloat32LE complex 32-bit floating-point (little-endian)
CFloat32BE complex 32-bit floating-point (big-endian)
CFloat64 complex 64-bit (double) floating-point (native endian-ness)
CFloat64LE complex 64-bit (double) floating-point (little-endian)
CFloat64BE complex 64-bit (double) floating-point (big-endian)

The image transfom

The orientation of the image with respect to the scanner axes is determined by the combination of the image axes and the location of the corner voxel. This information is encapsulated in the transformation matrix, commonly referred to simply as the transform. You can view the transform for any image using mrinfo, for example:

$ mrinfo dwi.mif
Image:               "dwi.mif"
  Dimensions:        104 x 104 x 54 x 167
  Voxel size:        2.30769 x 2.30769 x 2.3 x ?
  Data strides:      [ -1 -2 3 4 ]
  Format:            MRtrix
  Data type:         unsigned 16 bit integer (little endian)
  Intensity scaling: offset = 0, multiplier = 1
  Transform:               0.9999   6.887e-09    -0.01564      -116.1
                        -0.001242      0.9968    -0.07943      -89.44
                          0.01559     0.07944      0.9967      -64.27
  comments:          TOURNIER DONALD (BRI) [MR] diff60_b3000_2.3_iPat2+ADC
                     study: BRI_Temp_backup Donald
                     DOB: 09/03/1977
                     DOS: 03/10/2007 15:58:40
  dw_scheme:         [ 167 entries ]

The ‘transform’ field above shows the first 3 rows of the transformation matrix (technically, this is a 4×4 matrix, but the last row is always set to [ 0 0 0 1 ]). The first 3 columns correspond to the x, y & z image axes respectively, while the last column corresponds to the location in real (scanner/world) space of the corner voxel (i.e. the voxel at index [ 0 0 0 ]).

In MRtrix3, the transform shown always corresponds to the transformation from image coordinates in millimeters to scanner coordinates in millimeters - the voxel size is not taken into account, and the image axes are always normalised to unit amplitude. This may differ from other packages.

Furthermore, MRtrix3 will always present the transform that best matches the real space. If the transform of the image on file represents a large rotation, such that for example the image x axis is more closely with the scanner’s z axis, this transform will be modified by permutation or inversion of the axes to bring it in alignment with the expected coordinate system, so that the first axis genuinely can be interpreted as approximately left-right, etc. To achieve this, MRtrix3 will also modify the image Strides to match.


A file is simply a linear array of values. Image data on the other hand are multidimensional arrays. The image values can therefore be ordered on file in many different ways. For example, we could start from the voxel at the left posterior inferior corner of the image, and store intensity values in order of traversal towards the right. Once the other end of the image is reached, we repeat the process for the row of values anterior to the last one, and repeat until the end of the slice. At this point, we store the slice superior to the last one, until the whole image has been stored. This ordering scheme is what is typically used in the NIfTI standard, and is commonly referred to as RAS (right anterior posterior), referring to the direction of traversal of each axis in turn. This scheme is also often referred to as neurological, although this term is in general much more ambiguous.

However, this is only a convention, and many other combinations are possible. For instance, it is possible to start from the right posterior inferior corner, and raster through along the left direction, then store the next row along the anterior direction, and finally the next slice in the superior direction. This scheme is what is normally used in the now deprecated Analyse format, and is commonly referred to as RAS or radiological.

Of course, there are many more possibilities. For instance, sagittal DICOM images will typically be stored using a PIR (posterior inferior right) order, since each sagittal slice is stored in order, etc. MRtrix3 applications are agnostic to the order of storage, and can handle any such images provided the format is clear about what the order is.

In MRtrix3, the order of storage is defined by their strides. These refer to the number of voxels between a given voxel and the next voxel along a given dimension. For instance, in a 128×128×128 image stored using RAS ordering, the strides would be 1,128,16384: the next voxel along the x axis is just one voxel away, while the next voxel along the y axis is 128 values away (i.e. a whole row of x values), and so on. In contrast, if stored in LAS order, the strides would be -1,128,16384, indicating that the next voxel along the x axis would actually be stored one value before the current one.

To simplify the specification of these strides, MRtrix3 typically expects and provides symbolic strides. For example, the RAS strides above would be expressed as 1,2,3, since this is sufficient to deduce the actual strides once the image dimensions are known. Likewise, LAS would correspond to strides of -1,2,3, PIR to -2,-3,1, etc. This has the advantage that the specification of the strides is then independent of the image dimensions.

Using strides to specify ordering also allows the specification to generalise to arbitrary dimensions. For example, it is fairly common for MRtrix3 applications to request their output for 4D images to be written with strides 2,3,4,1 (if the image format supports it): this corresponds to a volume-contiguous order, whereby the values for all volumes of a given voxel are written next to each other on file; this often has performance advantages for applications that need to process all values for a given voxel concurrently (as is often the case in diffusion MRI), by allowing the hardware to make better use of resources (tractography is one such example).

Many MRtrix3 commands accept the -stride option, which is used to specify the strides for the output image. For example, to generate a LAS (radiological) NIfTI image for use with FSL (along with the corresponding bvecs/bvals), you can use mrconvert along with the -stride -1,2,3,4 option:

$ mrconvert dwi.mif -stride -1,2,3,4 -export_grad_fsl bvecs bvals dwi.nii

Likewise, if you need to ensure the orientation is neurological (RAS), you can specify strides 1,2,3,4 (or use the 1:4 shorthand). You can also specify other combinations if required: for example -stride -2,-1,3,4 would correspond to a PLS coordinate system, -stride 2,3,4,1 would correspond to volume-contiguous storage (with RAS for the spatial axes), etc.

The different formats supported by MRtrix3 differ in the range of strides that they support. The MRtrix image formats (.mih / .mif) are the only formats to support arbitrary combinations.


Not all image formats support all possible datatypes. The MRtrix image formats (.mih / .mif) are designed to handle arbitrary strides, while other image formats may only support a limited subset. When strides are requested that are not supported by the image format, a hopefully suitable alternative will be used instead.

Interaction between strides and transform

There is an interaction between the strides and the image transform: if the transform matrix corresponds to a 90° rotation, this can be viewed as changing the strides without affecting the transform. Such a large rotation has changed the order of storage relative to the anatomical labels typically used to refer to the ordering (e.g. RAS, LAS, etc). For example, if a RAS image is modified such that its transform rotates the image axes by 90° around the y axis, this in effect implies that voxels are now ordered IAR (i.e. right becomes inferior, anterior remains as-is, and superior becomes right).

The MRtrix3 back-end will indeed interpret such large rotations as affecting the strides, so that if the strides are stated as 1,2,3, the order of storage will always be left->right, posterior->anterior, inferior->superior relative to the scanner axes. Note that this also implies that the transform matrix will always be modified as necessary to bring it close to the standard coordinate system, so that the first image axis is close to the x axis, etc. This allows MRtrix3 applications to operate on images in the knowledge that these axes are always anatomically as expected, without worrying about the details of how this information was actually stored on file.

It is important to bear this in mind when interpreting for output of mrinfo for example, since this produces the strides and transform as interpreted by MRtrix3, rather than those actually stored on file - although the two representations should be strictly equivalent. If you need to inspect the information as stored on file, use mrinfo‘s -norealign option.

Supported formats

This lists the various file formats currently supported by MRtrix3.

MRtrix image formats (.mih / .mif)

These MRtrix-specific image formats are closely related. They consist of a text header, with data stored in binary format, either within the same file (.mif) or as one or more separate files (.mih). In both cases, the header structure is the same, as detailed below. These file formats were devised to address a number of limitations inherent in currently available formats. In particular:

  • simplicity: as detailed below, the header format is deliberately kept very simple and human-readable, making it easy to debug and edit manually if needed.
  • extendability: any information can be stored in the header, and will simply be ignored by the application if not recognised.
  • arbitrary data organisation: voxel values can be stored in any order, making it simple to ensure for example that all FOD coefficients for a given voxel are stored contiguously on file.

Note that MRtrix now includes MatLab functions to read and write MRtrix image files, and to load MRtrix tracks files. These are located in the matlab subfolder.

Compressed MRtrix image format (.mif.gz)

MRtrix3 also supports the compressed version of the single-file .mif format, both for reading and writing.


While this can reduce file sizes, it does incur a runtime cost when reading or writing the image (a process that can often take longer than the operation to be performed), and will require the entire image to be loaded uncompressed into RAM (MRtrix3 can otherwise make use of memory-mapping to keep RAM requirements to a minimum). For large files, these costs can become considerable; you may find that MRtrix3 can process a large uncompressed image, yet run out of RAM when presented with the equivalent compressed version (in such cases, you can try using gunzip to uncompress the file manually before invoking the relevant MRtrix3 command).

Header structure

The header is the first (and possibly only) data stored in the file, as ASCII-encoded text (although other encodings such as UTF8 may work equally well). Lines should be separated by Unix-style newlines (line-feed, ‘’, ASCII 0x0A), although MRtrix will also accept DOS-type newlines.

The first line should read only mrtrix image to indicate that this is an image in MRtrix format. The last line of the header should read only END to signal the end of the header, after which all data will be considered as binary.

All following lines are in the format key: value, with the value entry extending up to the end of the line. All whitespace characters before and after the value entry are ignored. Some keys are required to read the images, others are optional, and any key not recognised by MRtrix will simply be ignored. Recognised keys are listed below, along with the expected format of the corresponding values.

  • dim [required]

the image dimensions, supplied as a comma-separated list of integers. The number of entries specifies the dimensionality of the image. For example: dim: 192,256,256 specifies a 192×256×256 image.

  • vox [required]

    the voxel size along each dimension, as a comma-separated list of floating-point values. The number of entries should match that given in the dim entry. For example: vox: 0.9,0.898438,0.898438.

  • layout [required]

    specifies the organisation of the data on file. In simplest terms, it provides a way of specifying the strides required to navigate the data file, in combination with the dim entry. It is given as a comma-separated list of signed integers, with the sign providing the direction of data traversal with respect to voxel coordinates, and the value providing a way of specifying the order of increasing stride.

    For example, assuming an image with dim: 192,256,256, the entry layout: +2,-0,-1 is interpreted as: the shortest stride is along the y-axis (second entry), then the z-axis (third entry), and then along the x-axis. Voxels are stored in the order left to right (positive stride) along the x-axis; anterior to posterior along the y-axis (negative stride); and superior to inferior (negative stride) along the z-axis. Given the image dimensions, the final strides are therefore 256×256=65536 for adjacent voxels along the x-axis, -1 for the y-axis, and -256 for the z-axis. This also implies that the voxel at coordinate [ 0 0 0 ] is located 65536 voxel values into the data portion of the file.

  • datatype [required]

    the datatype used to store individual voxel values. See the listing of valid Data types. For example: datatype: UInt16LE

  • file [required]

    specifies where the binary image data are stored, in the format file: filename offset, with the offset provided in bytes from the beginning of the file. For example: file: image.dat 0.

    For the single-file format (.mif), the filename should consists of a single full-stop (‘.’) to indicate the current file, and the offset should correspond to a point in the file after the END statement of the header.

    For the separate header/data format (.mih), the filename should refer to an existing file in the same folder as the header (.mih) file. Multiple such entries can be supplied if the data are stored across several files.

  • transform [optional]

    used to supply the 4×4 transformation matrix specifying the orientation of the axes with respect to real space. This is supplied as a comma-separated list of floating-point values, and only the first 12 such values will be used to fill the first 3 rows of the transform matrix. Multiple such entries can be provided to fill the matrix; for example, MRtrix will normally produce 3 lines for the transform, with one row of 4 values per entry:

    transform: 0.997986,-0.0541156,-0.033109,-74.0329
    transform: 0.0540858,0.998535,-0.00179436,-100.645
    transform: 0.0331575,2.34007e-08,0.99945,-125.84
  • comments [optional]

    used to add generic comments to the header. Multiple such entries can be provided. For example: comment: some information

  • scaling [optional]

    used to specify how intensity values should be scaled, provided as an offset and scale. Voxel values will be read as value_returned = offset + scale * value_read. For example: scaling: -1,2. Default is 0,1 (no modification).

MRtrix sparse image formats (.msh / .msf)

These new image formats are designed for applications where the number of discrete elements within a voxel may vary between voxels. The most likely use case here is where each voxel contains some number of discrete fibre populations (‘fixels’), and some information associated with each of these elements must be stored. Since only as many elements are as required for any particular voxel are actually stored, rather than having to store the maximum possible number for all voxels and padding with empty data, the format is referred to as ‘sparse’.

Much like the standard MRtrix image formats (.mif and .mih), there are two different image file extensions available. One (.msh) separates the image header information and raw data into separate files, while the other (.msf) encodes all information relevant to the image into a single file.

However unlike these established formats, sparse images contain two separate raw data fields. The first of these behaves identically to standard images: a single intensity value for every image element. The second stores sparse image data. For any particular image element, the intensity value within the standard image field defines a pointer to a location within the sparse image field, where the sparse data relevant for that image element can be found.

Additional image header features

These image formats have some features within the image header that differ from the standard MRtrix image formats:

  • The ‘magic number’ that appears at the start of the file must read ‘mrtrix sparse image’.
  • Key:value pair ‘sparse_data_name’ defines the name of the class used in the sparse data field. This class name is typically not reader-friendly; the value that appears is that provided by the C++ call typeid(XYZ).name() for a class called XYZ. This is necessary to ensure that the data stored in the sparse field can be interpreted correctly.
  • Key:value pair ‘sparse_data_size’ defines the size (in bytes) of the class used to store the sparse data.
  • The ‘datatype’ field MUST be a 64-bit integer, with the same endianness as the system. A 64-bit integer type is required because the standard image data provides pointers to the sparse data in memory, while the endianness is tested to ensure that the sparse data can be interpreted correctly. Note that sparse images cannot be transferred and used between systems with different endianness.
  • In addition to the ‘file’ key, a second key ‘sparse_file’ is also required, which provides the path to the beginning of the sparse image data. In the .msf format, this provides an offset from the start of the file to the start of the sparse data field; in the .msh format, a second associated data file with the extension .sdat is generated on image creation, and the path to this file is defined in the header.

Sparse data storage

Within the sparse data field, there is no delimiting information or identifying features; the image format relies on the integers stored in the standard image field to provide offset pointers to appropriate locations within the sparse field.

From the data position defined by such an offset, the first 4 bytes provide a 32-bit integer (with native endianness), which specifies the number of discrete elements stored. This is followed by data to fill precisely that number of instances of the sparse data class. Note that no endianness conversion can be performed on this data; data is read and written using a straight memory copy.

DICOM (folder or .dcm)

DICOM format is only supported for reading. MRtrix3 applications will assume an image is in DICOM format if the image specifier provided corresponds to a folder or ends with the .dcm extension. For a folder, the application will scan the entire folder and its subfolders for DICOM files and generate a list of DICOM patients, studies and series. If a single series is found within the folder, this data set will be accessed with no further interaction required. Otherwise, the user will be prompted to select the series of interest. MRtrix3 supports data from all major manufacturers, including Siemens mosaics and the newer single-file multi-frame format.

A separate application, dcminfo, is provided to view all DICOM header elements within a particular DICOM file, including Siemens’ custom shadow attributes (CSA).

Note that no support is provided for reading the DICOMDIR entry due to case-sensitivity issues. DICOM data are typically stored on CD or DVD on a case-insensitive filesystem. However, Unix systems will typically not access these filesystems in a case-insensitive manner, and will fail to find the appropriate files if the case of filenames supplied in the DICOMDIR file does not match the case of the files found on the CD or DVD.

NIfTI (.nii)

This file format is supported both for reading and writing, and allows interoperation with other packages such as SPM or FSL.

if both qform and sform orientation fields are present, the qform fields are ignored. Obviously, the qform fields will be used if they are present on their own.

Compressed NIfTI (.nii.gz)

MRtrix3 also supports compressed NIfTI images both for reading and writing.


While this can reduce file sizes, it does incur a runtime cost when reading or writing the image (a process that can often take longer than the operation to be performed), and will require the entire image to be loaded uncompressed into RAM (MRtrix3 can otherwise make use of memory-mapping to keep RAM requirements to a minimum). For large files, these costs can become considerable; you may find that MRtrix3 can process a large uncompressed image, yet run out of RAM when presented with the equivalent compressed version (in such cases, you can try using gunzip to uncompress the file manually before invoking the relevant MRtrix3 command).

FreeSurfer formats (.mgh / .mgz)

MRtrix3 supports both of these formats for reading and writing.

Analyse format (.img / .hdr)

This file format is supported both for reading and writing. However, when writing, the newer NIfTI standard will be used, since the Analyse format cannot store crucial information such as the image transform, and is hence deprecated. If these images are actually stored as NIfTI, they will be handled appropriately according to the standard.


In order to specify an Analyse format image on the command line, type the name of its data file (*.img), not the header file.


By default, Analyse format images will be assumed to be stored using RAS (radiological) convention. This can modified in the `configuration`_ file, by setting the Analyse.LeftToRight entry to true.

Tracks file format (.tck)

The format for track files is similar to that for MRtrix image formats (.mih / .mif). It consists of a text header in the same key: value format, ending with a single ‘END’ statement, and followed by binary data.

The first line of the header should read mrtrix tracks to indicate that the file contains tracks in MRtrix format. Further key: value pairs typically provide information about the parameters used to produce the tracks, and for the most part are not required to read the data. The only required keys are the following:

  • file

    a file: . offset entry is required to specify the byte offset from the beginning of the file to the start of the binary track data. At this stage, only the single-file format is supported - in other words the filename part must be specified as ‘.’ (see above for details).

  • datatype

    specifies the datatype (and byte order). At this points only the Float32 data type is supported, either as little-endian (LE) or big-endian (BE).

    The binary track data themselves are stored as triplets of floating-point values (at this stage in 32 bit floating-point format), one per vertex along the track. Tracks are separated using a triplet of NaN values. Finally, a triplet of Inf values is used to indicate the end of the file.