You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Add channel_wise option to RandShiftIntensity and RandScaleIntensity #6629
It is very common practice to independently augment intensities of separate modalities. Yet currently RandShiftIntensity and RandScaleIntensity do not have this option, and instead randomly shift/scale along all channels at once.
It is noteworthy that RandStdShiftIntensity has a "channel_wise" option to do just this. I suggest we add a similar option to RandShiftIntensity and RandScaleIntensity.
Hi @philnovv, I think for RandScaleIntensity and RandShiftIntensity since we will randomly choose a factor or offset then no matter do it channel_wise or not, the result will be the same. Or do you mean randomly select a factor or offset for each channel?
If I have misunderstood, just correct me.
Thanks!
Hi @KumoLiu, the latter is what I mean. The option to randomly select and apply a factor/offset for each channel would be (I think) more closely aligned with standard practice.
BTW this applies to all intensity-based augmentation transforms. For example we may want to apply random blurring with an independent blurring kernel for each channel. This makes sense as an augmentation strategy because the different contrasts in real-world multi-contrast images are often acquired with very different resolutions/slice thicknesses etc...
Activity
KumoLiu commentedon Jun 28, 2023
Hi @philnovv, I think for
RandScaleIntensityandRandShiftIntensitysince we will randomly choose afactororoffsetthen no matter do it channel_wise or not, the result will be the same. Or do you mean randomly select afactororoffsetfor each channel?If I have misunderstood, just correct me.
Thanks!
philnovv commentedon Jul 19, 2023
Hi @KumoLiu, the latter is what I mean. The option to randomly select and apply a factor/offset for each channel would be (I think) more closely aligned with standard practice.
philnovv commentedon Jul 19, 2023
BTW this applies to all intensity-based augmentation transforms. For example we may want to apply random blurring with an independent blurring kernel for each channel. This makes sense as an augmentation strategy because the different contrasts in real-world multi-contrast images are often acquired with very different resolutions/slice thicknesses etc...
Add `channel_wise` in `RandScaleIntensity` (#6793)
fix Project-MONAI#6629
Add `channel_wise` in `RandShiftIntensity` (#7025)