Hamamatsu format

multi-file JPEG/NGR with proprietary metadata and index file formats, and single-file TIFF-like format with proprietary metadata
File extensions
.vms, .vmu, .ndpi
OpenSlide vendor backend


OpenSlide will detect a file as Hamamatsu if:

  1. The file given is a INI-style text file.
  2. It has a [Virtual Microscope Specimen] (VMS) or [Uncompressed Virtual Microscope Specimen] (VMU) group.
  3. If VMS, there are at least 1 row and 1 column of JPEG images (NoJpegColumns and NoJpegRows).

or if:

  1. The file has a TIFF directory structure.
  2. TIFF tag 65420 is present.


The Hamamatsu format has three variants. VMS and VMU consist of an index file, 2 or more image files, and (in the case of VMS) an “optimisation” file. NDPI consists of a single TIFF-like file with some custom TIFF tags. VMS and NDPI contain JPEG images; VMU contains NGR images (a custom uncompressed format).

Multiple focal planes are ignored, only focal plane 0 is read.

JPEG does not allow for files larger than 65535 pixels on a side. In VMS, multiple JPEG files are used to encode large images. To avoid having many files, VMS uses close to maximum size (65K by 65K) JPEG files. NDPI, instead, stuffs large levels into a single JPEG and sets the overflowed width/height fields to 0.

Unfortunately, JPEG provides very poor support for random-access decoding of parts of a file. To get around this, JPEG restart markers are placed at regular intervals, and these offsets are specified in the optimisation file (in VMS) or in a TIFF tag (in NDPI). With restart markers identified, OpenSlide can treat JPEG as a tiled format, where the height is the height of an MCU row, and the width is the number of MCUs per row divided by the restart marker interval times the width of an MCU. (This often leads to oddly-shaped and inefficient tiles of 4096x8, for example.)

Unfortunately, the VMS optimisation file does not give the location of every restart marker, only the ones found at the beginning of an MCU row. It also seems that the file ends early, and does not give the location of the restart marker at the last MCU row of the last image file.

Thus, the optimisation file can only be taken as a hint, and cannot be trusted. The entire set of JPEG files must be scanned for restart markers in order to facilitate random access. OpenSlide does this lazily as needed, and also in a background thread that runs only when OpenSlide is otherwise idle.

The VMS map file is a lower-resolution version of the other images, and can be used to make a 2-level JPEG pyramid. JPEG also allows for lower-resolution decoding, so further pyramid levels are synthesized from each JPEG file.

VMS File

The .vms file is the main index file for the VMS format. It is a Windows INI-style key-value pair file, with sections. Only keys in the Virtual Microscope Specimen group are read by OpenSlide.

Here are known keys from the file:

Key Description
NoLayers Number of layers, currently must be 1 to be accepted
NoJpegColumns Number of JPEG files across, given in ImageFile attributes
NoJpegRows Number of JPEG files down, given in ImageFile attributes
ImageFile Semantically equivalent to ImageFile(0,0,0), though not specified that way. The image in position (0,0,0) of the set of images
ImageFile(x,y) Semantically equivalent to ImageFile(0,x,y), though not specified that way. The image in position (0,x,y) of the set of images
ImageFile(z,x,y) Where x and y are non-negative integers. Both x and y cannot be 0. z is a positive integer. These are the images that make up the virtual slide, as a concatenation of JPEG images. x and y specify the location of each JPEG, z specifies the focal plane
MapFile A lower-resolution version of all the ImageFiles
OptimisationFile File specifying some of the restart marker offsets in each ImageFile
AuthCode Unknown
SourceLens Apparently the objective power
PhysicalWidth Width of the main image in nm
PhysicalHeight Height of the main image in nm
LayerSpacing Unknown
MacroImage Image file for the “macro” associated image
PhysicalMacroWidth Width of the macro image in nm
PhysicalMacroHeight Height of the macro image in nm, sometimes with a trailing semicolon
XOffsetFromSlideCentre Distance in X from the center of the entire slide (i.e., the macro image) to the center of the main image, in nm
YOffsetFromSlideCentre Distance in Y from the center of the entire slide to the center of the main image, in nm

VMU File

The .vmu file is the main index file for the VMU format. Only keys in the Uncompressed Virtual Microscope Specimen group are read by OpenSlide.

Here are known keys from the file:

Key Description
NoLayers (see VMS above)
ImageFile (see VMS above)
ImageFile(x,y) (see VMS above)
ImageFile(z,x,y) (see VMS above)
MapFile (see VMS above)
MapScale Seems to be the downsample factor of the map
AuthCode (see VMS above)
SourceLens (see VMS above)
PixelWidth Width of the image in pixels
PixelHeight Height of the image in pixels
PhysicalWidth (see VMS above)
PhysicalHeight (see VMS above)
LayerSpacing (see VMS above)
LayerOffset Unknown
MacroImage (see VMS above)
PhysicalMacroWidth (see VMS above)
PhysicalMacroHeight (see VMS above)
XOffsetFromSlideCentre (see VMS above)
YOffsetFromSlideCentre (see VMS above)
Reference Unknown
BitsPerPixel Bits per pixel, currently expected to be 36
PixelOrder Currently expected to be RGB
Creator String describing the software creating this image
IlluminationMode Unknown
ExposureMultiplier Unknown, possibly the multiplier used to scale to 15 bits?
GainRed Unknown
GainGreen Unknown
GainBlue Unknown
FocalPlaneTolerance Unknown
NMP Unknown
MacroIllumination Unknown
FocusOffset Unknown
RefocusInterval Unknown
CubeName Unknown
HardwareModel Name of the hardware
HardwareSerial Serial number of the hardware
NoFocusPoints Unknown
FocusPoint0X Unknown
FocusPoint0Y Unknown
FocusPoint0Z Unknown
FocusPoint1X Unknown
FocusPoint1Y Unknown
FocusPoint1Z Unknown
FocusPoint2X Unknown
FocusPoint2Y Unknown
FocusPoint2Z Unknown
FocusPoint3X Unknown
FocusPoint3Y Unknown
FocusPoint3Z Unknown
NoBlobPoints Unknown
BlobPoint0Blob Unknown
BlobPoint0FocusPoint Unknown
BlobPoint1Blob Unknown
BlobPoint1FocusPoint Unknown
BlobPoint2Blob Unknown
BlobPoint2FocusPoint Unknown
BlobPoint3Blob Unknown
BlobPoint3FocusPoint Unknown
BlobMapWidth Unknown
BlobMapHeight Unknown


NDPI uses a TIFF-like structure, but libtiff cannot read the headers of an NDPI file. This is because NDPI specifies the RowsPerStrip as the height of the file, and after doing out the multiplication, this typically overflows libtiff and it refuses to open the file. Also, the TIFF tags are not stored in sorted order.

NDPI stores an image pyramid in TIFF directory entries. In some files, the lower-resolution pyramid levels contain no restart markers. The macro image, and sometimes an active-region map, seems to come last.

JPEG files in NDPI are not necessarily valid. If ImageWidth or ImageHeight exceeds the JPEG limit of 65535, then the width or height as stored in the JPEG file is 0. libjpeg will refuse to read the header of such a file, so the JPEG data stream must be altered when fed into libjpeg.

NDPI is based on the classic TIFF format, which does not support files larger than 4 GB. However, NDPI files can be larger than 4 GB. NDPI generally handles this by overflowing the corresponding TIFF fields, requiring the reader to guess the high-order bits. This affects TIFF Value Offsets with pointers to out-of-line values, as well as the value of the StripOffsets field. Some TIFF fields (e.g. Software) have the same Value Offset in every directory; for these, no concatenation of high-order bits is necessary. For the others (primarily field 65426) it seems reasonable to select high-order bits which place the value at the largest offset below the directory itself, since the TIFF directory is positioned after the data it points to. NDPI always stores next-directory offsets (in the TIFF header and at the end of each directory) as 64-bit quantities, even though TIFF specifies them as 32 bits; this is possible because the TIFF standard places them at the end of their parent data structures.

It is not clear whether NDPI can support individual directories larger than 4 GB. Such files would require additional inferences for the StripByteCounts field, for Value Offsets that are identical across directories, and for the optimisation entries.

Here are the observed TIFF tags:

Tag Description
ImageWidth Width of the image
ImageHeight Height of the image
Make “Hamamatsu”
Model “NanoZoomer” or “C9600-12”, etc
XResolution Seemingly correct X resolution, when interpreted with ResolutionUnit
YResolution Seemingly correct Y resolution, when interpreted with ResolutionUnit
ResolutionUnit Seemingly correct resolution unit
Software “NDP.scan”, sometimes with a version number
StripOffsets The offset of the JPEG file for this layer
StripByteCounts The length of the JPEG file for this layer
65420 Always exists, always 1. File format version?
65421 SourceLens, correctly downsampled for each entry. -1 for macro image, -2 for a map of non-empty regions.
65422 XOffsetFromSlideCentre
65423 YOffsetFromSlideCentre
65424 Seemingly the Z offset from the center focal plane (in nm?)
65425 Unknown, always 0?
65426 Optimisation entries, as above
65427 Reference
65428 Unknown, AuthCode?
65430 Unknown, have seen 0.0
65433 Unknown, I have seen 1500 in this tag
65439 Unknown, perhaps some polygon ROI?
65440 Unknown, I have seen this: <0 0 0 1 0 2 0 3 0 4 0 5 0 6 0 7 0 8 1 9 1 10 1 11 1 12 1 13 1 14 1 15 1 16 1 17>
65441 Unknown, always 0?
65442 Scanner serial number
65443 Unknown, have seen 0 or 16
65444 Unknown, always 80?
65445 Unknown, have seen 0, 2, 10
65446 Unknown, always 0?
65449 ASCII metadata block, key=value pairs, not always present
65455 Unknown, have seen 13
65456 Unknown, have seen 101
65457 Unknown, always 0?
65458 Unknown, always 0?

Optimisation File (only for VMS)

The optimisation file contains a list of 32- (or 64- or 320- ?) bit little endian values, giving the file offset into an MCU row, each offset starts at a 40-byte alignment, and the last row (of the entire file, not each image) seems to be missing. The offsets are all packed into 1 file, even with multiple images. The order of images is left-to-right, top-to-bottom.

Map File (only for VMS/VMU)

The VMS map file is a standard JPEG file. Its restart markers (if any) are not included in the optimisation file. The VMU map file is in NGR format. This file can be used to provide a lower-resolution view of the slide.

Image Files (only for VMS/VMU)

These files are given by the VMS/VMU ImageFile keys. They are assumed to have a height which is a multiple of the MCU height. They are assumed to have a width which is a multiple of MCUs per row divided by the restart interval.

For VMS, these files are in JPEG, for VMU they are in NGR format.

NGR Format

The NGR file contains uncompressed 16-bit RGB data, with a small header. The files we have encountered start with GN, two more bytes, and then width, height, and column width in little endian 32-bit format. The column width must divide evenly into the width. Column width is important, since NGR files are generated in columns, where the first column comes first in the file, followed by subsequent files. Columns are painted left-to-right.

At offset 24 is another 32-bit integer which gives the offset in the file to the start of the image data. The image data we have encountered is in 16-bit little endian format.

Associated Images

the image file given by the MacroImage value in the VMS/VMU file, or SourceLens of -1 in NDPI

Known Properties

All key-value data stored in the VMS/VMU file, and known tags from the NDPI file, are encoded as properties prefixed with “hamamatsu.”.

For VMS, calculated as hamamatsu.PhysicalWidth/(1000*openslide.level[0].width). For NDPI, calculated as 10000/tiff.XResolution, if tiff.ResolutionUnit is centimeter.
For VMS, calculated as hamamatsu.PhysicalHeight/(1000*openslide.level[0].height). For NDPI, calculated as 10000/tiff.YResolution, if tiff.ResolutionUnit is centimeter.
normalized hamamatsu.SourceLens

Test Data

NDPI format
VMS format