I am beginning to use QSIPrep for my research and am very impressed with its capabilities. I have a couple of questions regarding the recommended denoising workflow that I hope you can clarify.
My main question concerns the academic acceptability of using Patch2Self as the primary denoising method for a publication. My literature search indicates that while many papers compare Patch2Self to other methods, I have not found many applied group-level studies that use it as their chosen preprocessing step. This has made me wonder if there are any potential concerns or limitations regarding its use that I should be aware of when preparing a manuscript.
Secondly, I’ve noticed that the QSIPrep pipeline performs denoising (such as Marchenko-Pastur Principal Component Analysis, or MP-PCA, and Patch2Self) before the eddy correction step. Could you please elaborate on the rationale for this particular order? Are there any circumstances where you would recommend reversing this and running eddy correction before denoising?
Thank you for your time and for developing this valuable tool. Any insights you could provide would be greatly appreciated.
Empirically, we have noticed that motion estimates are not as accurate on data denoised with Patch2Self. Not sure how ubiquitous that is though, and may not have a big deal if motion is not a covariate in your model.
Techniques like MP-PCA rely on accurate modeling of the noise covariance structure in the raw k-space-derived image. Once motion or eddy-current corrections are applied (which involve interpolation and resampling), the original noise structure is no longer preserved.
MP-PCA and Patch2Self work best on raw data because these methods assume voxel-wise independence in their noise models—an assumption violated after spatial transformations.
Denoised data has a higher signal, which can improve the accuracy of image registration algorithms used in eddy correction.
Thank you so much for your reply; it was very easy to follow. My first question is about using an AI method that could be seen as a “black box.” Although I am persuaded by Patch2Self, particularly because it is a self-supervised algorithm, I’m not sure how others (for example, a reviewer) might react to me using it as a denoiser. I’m concerned they might criticize this choice.