Version 1.9.0

• Methods for the Math, Ops, Complex and Summary group generics have
  been added for the internalImage class, as well as is.na() and mean().
  This prevents R from falling back on the character methods, which aren't
  helpful. In most cases these implementations simply call the relevant generic
  on the array version of the image; the exception is Summary, which calls a
  C++ implementation to avoid allocating memory for an R array when only a
  summary value is needed. (Reported by Martin Wilson, issue #39 and divest
  issue #6.)
• Image value range calculation and NA handling in the core C++ code has been
  improved, particularly for the 32-bit integer datatype.

Version 1.8.0

• It is now possible to read, write and manipulate additional image metadata
  fields stored in JSON "sidecar" files like those used by the Brain Imaging
  Data Structure. readNifti and writeNifti each gain a "json" argument,
  and the new imageAttributes function can be used to get and set attributes
  for images. Conversion to or from JSON requires the jsonlite package;
  this is an R-only feature for now.
• If a plain string is passed to lyr as the "scale" argument and doesn't
  match a known scale name from the shades package, it is now assumed to be
  a single literal colour name. (Prompted by Ellyn Butler, issue #36.)
• Initialisation of the complex NA value in C++ code has been simplified to
  avoid remaining compilation problems on some platforms. (Issues #35 and #38.)

Version 1.7.0

• Indexing into an rgbArray or an RGB-valued internalImage now produces a
  result of type rgbArray, to maintain the special interpretation of these
  values, and facilitate conversion to strings.
• rgbArray now also gains an as.raster method.
• The test suite has been moved from testthat to tinytest.
• An API accommodation for versions of R before 3.1.0 has now been removed, as
  it produces a note in the current R-devel.

Version 1.6.1

• A misaligned memory read caught by clang-UBSan has been corrected.

Version 1.6.0

R interface

• The niftiHeader() function gains an "unused" argument. If TRUE, the
  function's return value includes those legacy ANALYZE-7.5 fields that are
  formally unused by the NIfTI-1 standard (or the last few bytes of a NIfTI-2
  header). These fields cannot be set, however, as the niftilib I/O functions
  discard them, and the core nifti_image structure does not represent them.
  They also cannot be accessed with the $ operator shorthand. The option is
  therefore only useful to inspect file headers created by other software.

API changes

• The NiftiImage class now has constructors taking a NIfTI header structure.

Other changes

• The NiftiImageData::Iterator inner class now holds a pointer, not a
  reference, to its parent. This doesn't really change the API, but ensures
  that iterator objects are mutable, which is formally required by the C++
  standard. (Prompted by Timothy Anderson, issue #31.)
• There is now support for standalone use on Windows with the Visual C++
  toolchain, including an nmake-compatible makefile, and this is now tested
  as part of package CI. The package also no longer includes symlinks, for
  better portability. (Prompted by Timothy Anderson, issue #31.)
• The bundled zlib is now version 1.3. The included zlib headers are no
  longer exposed to client packages if the system zlib is used for linkage,
  as the two may be incompatible.

Version 1.5.1

• Several warnings from recent compilers relating to printf-style format
  strings have been resolved.

Version 1.5.0

• The writeNifti() and writeAnalyze() functions gain a "compression"
  argument to control the compression level used when writing gzipped images.
• The viewer will now show 4D images with fourth dimension 3 as vectors if the
  RNifti.d4vectors option is TRUE. This is off by default (following the
  previous behaviour) as it's less explicit than an image with a vector intent.
• The niftiHeader() function will no longer call asNifti() on an argument
  that looks like a path, as this will wastefully read in the pixel data when
  only the metadata is needed.
• Support for MriImage objects with complex or RGB types has been added.
• cfloat and cdouble are now additionally accepted as datatypes for 32-bit
  and 64-bit floating-point complex types, respectively.
• The upstream NIfTI libraries have been updated.