Question
Question Posted 04/09/26:
Dear ADNI Experts,
I am working on a deep learning project for Alzheimer’s disease diagnosis using ADNI MRI data. I would like to ask whether differences in ADNI MRI preprocessing labels, such as GradWarp and N3/N3m, could substantially affect downstream image classification results.
In my dataset, some T1 MRI scans are labeled as “MPR; GradWarp”, while others are labeled as “MT1; N3m”. My concern is whether these images are significantly different in a way that could bias or affect a deep learning model for AD classification.
More specifically, I would like to ask:
Are images with labels such as MPR; GradWarp and MT1; N3m considered substantially different for imaging analysis?
After standard preprocessing on my side, would these differences still be expected to have a large impact on classification performance?
Is it acceptable to use these image types together in one dataset, or would you recommend restricting the analysis to one more homogeneous image type?
If mixing them is unavoidable due to limited sample size, is there a preferred choice or recommended strategy?
I would greatly appreciate any guidance on whether these preprocessing differences are likely to be a major issue for deep learning-based AD diagnosis.
Thank you very much for your help.
Dear ADNI Experts,
I am working on a deep learning project for Alzheimer’s disease diagnosis using ADNI MRI data. I would like to ask whether differences in ADNI MRI preprocessing labels, such as GradWarp and N3/N3m, could substantially affect downstream image classification results.
In my dataset, some T1 MRI scans are labeled as “MPR; GradWarp”, while others are labeled as “MT1; N3m”. My concern is whether these images are significantly different in a way that could bias or affect a deep learning model for AD classification.
More specifically, I would like to ask:
Are images with labels such as MPR; GradWarp and MT1; N3m considered substantially different for imaging analysis?
After standard preprocessing on my side, would these differences still be expected to have a large impact on classification performance?
Is it acceptable to use these image types together in one dataset, or would you recommend restricting the analysis to one more homogeneous image type?
If mixing them is unavoidable due to limited sample size, is there a preferred choice or recommended strategy?
I would greatly appreciate any guidance on whether these preprocessing differences are likely to be a major issue for deep learning-based AD diagnosis.
Thank you very much for your help.
Response posted 04/09/26 by Jeff:
It depends. Are you expecting longitudinal consistency? GradWarp is, as the name hints, correcting a spatial distortion in the image due to imaging gradient hardware limitations. The distortion is fixed in space in the scanner, patients are not fixed in the scanner from imaging study to imaging study. The result being that relative to the anatomy the distortion appears to vary over time. Here's where it gets tricky: some scanners did 3D distortion correction on the scanner; some did 2D, and some did none. GW files should have been corrected to 3D in post-processing. If no GW version exists for a particular T1, then it was done in 3D on the scanner. In ADNI and ADNI-2 it's a mix. By ADNI-3 all scanners should have been doing 3D correction on the scanner. N3 attempts to correct for shading artifacts due to RF non-uniformity in the head coil. Whether that matters depends on the assumptions in your model.
