How does torchio.transforms.RescaleIntensity's in_min_max argument works?

[gRox167]

torchio.transforms.RescaleIntensity's in_min_max argument is not working properly or maybe I misunderstood the document.

Expected behavior
In the document, it is said that "Range (mmin,mmax) of input intensities that will be mapped to (nmin,nmax). If None, the minimum and maximum input intensities will be used." I thought it works like the function torch.clamp or np.clip, which will clip the in_min and in_max.
Actual behavior
In your source code, this step is done by

array -= in_min
in_range = in_max - in_min
array /= in_range

I think this just scale down the distribution of image intensity, but not clip off the value outside the range we set.

[romainVala]

Hello,

no there is a clip first : https://github.com/fepegar/torchio/blob/c95410bc1eeb8066907a3ba0eb152d75afdffd7d/torchio/transforms/preprocessing/intensity/rescale.py#L81
and then the rescale

[fepegar]

Hi, @gRox167. I'm not sure what the question is.

The clipping happens only if the percentiles kwarg is used (to remove potential outliers and make the transform more robust). Otherwise there's no clipping, just a linear transform to change the range of intensity values. You can take a look at the scikit-learn docs for rescale_intensity, it might help.

[gRox167]

Thanks for your reply! But I'm still a little confused about the behavior of in_min_max in Torchio, for it's different from scikit-learn.
In rescale_intensity the document mentioned that if the min/max value of in_range is more/less than the min/max image intensity, then the intensity levels are clipped

>>>image = np.array([51, 102, 153], dtype=np.uint8)
>>>image_float = 1.0 * image
>>>rescale_intensity(image_float, in_range=(0, 102))
array([0.5, 1. , 1. ])

The expected behavior of in_min_max is to clip off the outlier that smaller or larger than in_min or in_max, as mentioned in scikit-learn's rescale_intensity. So I wonder maybe the document is a little confuse for the novice user :), as it works differently from scikit's rescale_intensity.

[fepegar]

Well, I suppose the implementation in TorchIO is different, then. Do you think the skimage one makes more sense?

The expected behavior of in_min_max is to clip off the outlier that smaller or larger than in_min or in_max, as mentioned in scikit-learn's rescale_intensity

That's not really true, the expected behaviour is not to clip. That's what percentiles does.

Let's see if I can upload a photo from my phone...

[fepegar]

So here in1, in2 = in_min_max and out1, out2 = out_min_max.

Before that linear mapping, the array is clipped using the given percentiles to remove outliers.

[gRox167]

I guess I get it. But in TorchIO's code, the in_min_max only clip off the lower boundary limited by in_min and divide all intensity by in_max-in_min. Is there an inconsistency in its logic, why not clip off the upper bound like scikit? I know percentiles done the clip-off work, and in_min_max also do clip-off work in scikit style rescale_intensity, these 2 clip off are not compatible. So I guess it's better to let users choose which way to go.

And also your ways of implementation will cause the output result doesn't always fall into the range of (out_min,out_max).

array -= in_min
in_range = in_max - in_min
array /= in_range

In this code snippet, if in_max is smaller than the original max intensity, the array's range is not (0,1), max value will larger than 1, which will lead to a larger output value in the later step.

Also in the photo, if the whole x-axis represents the total intensity distribution such as (0,255), the $in_1$, $in_2$ limited the input of intensity to such as (50,200), then rescale voxel intensity between (50,200) to $out_1$, $out_2$. Does the photo represent the same behavior as a scikit style rescale_intensity.

Thanks for your kindly answer.