Closed vkt1414 closed 10 months ago
Just one small fix is needed, unless I messed it up - looks good otherwise!
Thank you! I corrected the units now.
To follow up on the above, it seems that pyradiomics will always calculate statistical minimum/maximum (see https://github.com/AIM-Harvard/pyradiomics/blob/master/radiomics/firstorder.py#L71-L95), without binning. Which means we should use IBSI code 1GSF
(see below).
However, the standard appears to only contain codes for discretized intensity 1PR8
:
See in part 16: https://dicom.nema.org/medical/dicom/current/output/chtml/part16/sect_cid_7478.html.
@dclunie why is the first type of intensity feature with code 1GSF
is not included in the standard?
@vkt1414 we discussed the above with @dclunie and confirmed we should use the appropriate code from IBSI even if it is not included in the standard. David will amend the standard to include missing codes.
@dclunie to follow up on what you said: what are the DICOM means to know the input units of pixel intensity other than Modality for CT?
Per discussion, for now we will abort the workflow if the input is not CT, but for the reference, here are some comments from @dclunie:
RescaleType
https://dicom.innolitics.com/ciods/ct-image/ct-image/00281054 is intended for communicating units, but looking at our data, it often contains invalid informationDICOM CT images are HU unless otherwise specified:
and for multi-energy CT it has become complicated:
https://dicom.nema.org/medical/dicom/current/output/chtml/part03/sect_C.11.html#sect_C.11.1.1.2
There is also PET, which has an attribute called Units:
https://dicom.nema.org/medical/dicom/current/output/chtml/part03/sect_C.8.9.html#sect_C.8.9.1.1.3
This PR addresses #33 and thanks for catching the error.
Summary: Added
Hounsfield Unit
as the unit for all intensity fields based on IBSI documentation.Here was my understanding upon reading IBSI paper. https://pubs.rsna.org/doi/suppl/10.1148/radiol.2020191145/suppl_file/IBSI_Reference_Manual.pdf
We must use calibrated units for CT and for modalities like MRI we must use arbitrary units a/c below paragraph from IBSI. Examples of calibrated units were Hounsfield for CT and SUV for PET
That means we must use
Intensity-based statistical features
to encode all intensity-based radiomics features for modalities like CT andIntensity histogram features
for modalities like MRI.However, there are two features that pyradomics refers to Intensity histogram features by default,
For units, this was my thought process, Wherever
Xgl
is present, I assumed the unit will beHounsfield Unit
and X^2 will besquare Hounsfield Unit
. While I could findHounsfield Unit
as [hnsf'U], I presumed,square Hounsfield Unit
as [hnsf'U]2 but we may need to confirm it with D.Clunie.It would be great to have a second set of eyes to confirm that the units are correct, as this seems to be vulnerable to errors.