How are these normalized at the point of the scanner? Are these scaled, or more related to actual physical units?
Created by Clinton Mielke subcosmos Actually the xray dose, exposure times, and other physical parameters are in the dicom header. My model takes these into account, but Im unable to test it since there's little room to submit jobs on the system.
Im bummed. Lots of good ideas with no time to test. Usually the raw image is generated by the mammography unit in relation with the xray dose absorbed by the receptor for each pixel on the receptor without scaling. There is usually a physical filter inside the machine to lower the scatter noise from xray dispersion; this is applied to optimize the signal to noise ratio for a specific xray dose.
Then, on a software basis, we can usually choose the default image filter, between different image filters (smooth, sharp, ...), applied by the unit to the raw data before sending the dicom image to the PACS (archiving) system. This default image filter, usually chosen by the local radiologists, can be different for each manufacturer and different between clinics/hospitals using the same machine.
With a sharp filter, the image is more sharp (of course!) but more noisy, with a default mean value probably higher than the raw image.
With a smooth filter, the image is more smooth(surprise!) but less noisy, with a default mean value probably lower than the raw image.
If we had access to the xray dose in the dicom image, a cnn model could probably retrospectively detect the filter applied after some supervised training.
For a specific radiologist, this filter can potentially change the sensibility and specificity of the interpretation according to his own experience. In practice, with a training database of this size (640000 images), a decent CNN deep model architecture probably automatically catches and adapts to this variation of filter.
Alexandre.
I'm not sure if the specifics of how each manufacturer does it, but the output values should correspond to the actual density of the breast.
All machines at accredited institutions in the US are tested periodically (http://www.acraccreditation.org/~/media/ACRAccreditation/Documents/Resources/DMQC/Mammo-QC-Test-Forms_Technologists_Excel.xlsx?la=en) and specifically for image quality they are all tested against the same phantom with multiple fibers, masses, and clusters of varying sizes and densities.
To pass, a radiologist needs to be able to discern 3/5 masses, 3/5 clusters, and 4/6 fibers on the phantom.
It might be interesting to see what our models would make of a phantom QC image. Hi Clinton,
I guess that the pixel values are scaled anyway to fit into the bit depth selected by the manufacturer of the scanner. I don't know much about how the different scanners generate the pixel values.