NumPy Rendered Media Type in DICOM

[Brian Bialecki]

Hello,

I am working on a work item proposal with DICOM WG27 to allow for requests of image/numpy and series/numpy as rendered media types.  I wish to have this added to the DICOM standard to ensure that arrays are produced that can be easily understood and there would be consistency in they way they are produced.  I have several items I know must be specified and here are some ideas I have to create consistency:

·         Pixel Data must be decompressed and represented as Little Endian

·         All Photometric Interpretation will be represented as Monochrome2

·         All Pixel Data will be represented as an unsigned integer

·         For Color Pixels should be always represented color by pixel, R1, G1, B1, R2, G2, B2

·         All color representations converted to RGB per Part 3 section C.7.6.3 where then shape will determine color from monochrome

·         Error should be thrown for images with lossy compression that data cannot be retrieved in this format for example urn:dicom:wado:0007 - The requested instance(s) cannot be provided in the requested format or transfer syntax.

·         For 3D arrays pixel spacing and slice thickness must be the same for all slices or frames, values returned in header

Open issues to discuss are:

·         How will the conversion to unsigned integers be handled?

·         Do we use raw pixel data representations only or pixel data rendered with machine applied LUTS?

·         How will pixel padding be defined?

I was hoping for some feedback from this group on items you may want to see, issues that you have encountered, or other best practices that I may need to ensure are addressed.

I am thinking something of a structure as follows:

Headers:

 

ndim

number of dimensions, 3 for 2D monochrome or RGB, 4 for 3D (multi-frame or series) monochrome or RGB

 

shape

2D image shape = (0028,0010) Rows, (0028,0011) Columns, (1=monochrome or 3=RGB)

Multi Frame = Frames, (0028,0010) Rows, (0028,0011) Columns, (1=monochrome or 3=RGB)

Series = Instances, (0028,0010) Rows, (0028,0011) Columns, (1=monochrome or 3=RGB)

 

size

total number of elements in the array

 

dtype

int16 – from bits stored (0028,0101), may be int8, int12 or int16 for example

 

itemsize

the size (in bytes) of each element of a NumPy array

 

To represent pixel spacing there needs to be an additional header field to represent

Pixel Spacing (0028,0030)

Slice Thickness (0018,0050)

Any feedback is welcome.