We run presurgical language mapping and have recently moved from a Philips scanner to a United Imaging (both 3 tesla new scanner is PET MRI). We scanned the same patient on both systems using the same task, and I would appreciate guidance on whether a difference we are seeing is a real problem or only a difference in vendor scaling.
The mean raw EPI intensity across a run is approximately 1300 on the Philips data and approximately 450 on the United Imaging data. Following the vendor's recommendation we increased TE from 24.9 ms to 30 ms, which raised TR accordingly, and the mean is now closer to 400.
My concern came from an AFNI Bootcamp discussion of losing numerical accuracy when working with small-valued integer data. My question is whether a mean of roughly 400 is low enough for quantization to matter in practice.
My own reasoning is that it should not be, because the quantization step of 1 unit is far below the temporal standard deviation, so noise dithers the signal across many levels. But I would rather confirm that than assume it.
Acquisition details:
United Imaging uPMR 790, epi_bold
36 slices, 3 mm thick, 2 x 2 mm in-plane, 136 volumes
Converted with Dimon, analysis in native (+orig) space
Measured values inside a brain mask:
Philips: mean , temporal SD , TSNR
United Imaging: mean , temporal SD , TSNR
Specific questions:
Given the temporal SD above, is quantization a genuine concern here, or is TSNR the only comparison that matters?
Is there anything in the AFNI conversion path (Dimon or to3d) that could be discarding a DICOM rescale slope, or does a low mean simply reflect how this vendor scales its reconstruction?
Would converting to float early in the pipeline help, or is the precision already fixed at the DICOM level such that this only adds file size?
Is there a recommended AFNI-side check for comparing data quality across vendors beyond TSNR?
Indeed, 400-450 is low and will cause numerical truncation akin to adding noise to a time series. To be sure, are those means computed over brain-masked voxels? (oh, you already answered that they are, thanks)
It is very likely that you can control that aspect of the scaling on the scanner, perhaps when you create a protocol (though I am not a MRI physicist, and my expectations might not match reality).
If you find that you are able to control this, definitely get it over 1000. With a mean of 500, a change by 1 is a change by 0.2%, which is the smallest change the scanner can record. If your experiments look at contrasts, or worse, interactions, and especially if events are short, those betas might drop well below this noise floor. Then the results might be much harder to detect.
Some scanners use 12 bits to record data, I don't know what the United Imaging scanners use. But that would mean a maximum (not mean) of 4095 in data values. So having a mean around 1500 might make sense in such a case (providing bright voxels are not hitting that 4095 limit).
Anyway, find out if you can control the scaling, and find out how many bits are actually used for storing the signal.
I spoke with the United Imaging scientist, and they informed me that the scanner stores the data using 16 bits. I also checked my current protocol, and the scaling factor (CBSF) is set to 1.
Do you think increasing the scaling factor to 20 would be helpful for improving the EPI signal and reducing the numerical truncation issue you mentioned?
It might be good to get a histogram of the current values. With a mean in the 400-500 range, the upper values presumably exceed 1000. I would be inclined to scale that up by about 5. The important thing is to get the output data in a reasonable range. Do not use all 16 bits either, for example, which applies to the maximum and not the mean. It is preferable for the maximum to fit comfortably as well.
From what I know, the magnitude of the data is based on the scanner protocol, applied in the reconstruction software. Maybe the United Imaging scientist has an idea of the expected range of EPI values when CBSF=1, and from there an alternative scale factor might be more clear. Or maybe that is not actually known on their end, but hopefully they can advise better than I.
Note that this will vary across subjects and how they are physically in the scanner. So it is good to have wiggle room on either end.
Hopefully it will not be necessary for the value to be adjusted every time. Because someone will forget, and losing data would be frustrating.
-rick
The
National Institute of Mental Health (NIMH) is part of the National Institutes of
Health (NIH), a component of the U.S. Department of Health and Human
Services.