How to properly rescale pixel values to be in the range [0, 1]?

[CristianoPizzamiglio]

Thanks for this beautiful library!

This question is very similar to the one asked on SO. However, I am struggling to apply the suggested solution.

I have loaded a CT dicom image having 512 rows and 512 columns:

dataset = pydicom.dcmread(dicom_path)

Then, I have created a 2D numpy array, called mask, having shape (512, 512) that consists only of 0s and 1s.

I have converted mask values to be np.float16 as suggested here and then I have overwritten PixelDataas explained here:

mask = mask.astype(np.float16)
dataset.PixelData = mask.astype(np.float16).tobytes()

Finally, I have written the modified dicom to file:

dataset.save_as("dicom_modified.dcm")

However, when I open the newly created file with a DICOM viewer software, instead of having 0s and 1s as pixel values, I get -1024s and 2048s.

My first solution was to applied the method proposed in the comment by @scaramallion in the SO link provided above:

if dataset.PixelRepresentation == 0:
    mask = mask / (dataset.BitsStored**2 - 1)
else:
    mask = (mask + (dataset.BitsStored - 1)**2) / (dataset.BitsStored**2 - 1)

but now I get -1024s and 2857s as pixel values. Maybe I have not properly understood @scaramallion comment and so the formulas are not correct.

Lastly, I have modified RescaleIntercept and RescaleSlope and set them to 0 and 1 / 2048 respectively. Now I am actually getting 0s and 1s in the newly created dicom file. But considering the answer by @scaramallion in the SO link above, I wonder if this is the proper way to normalize the pixel values.

Thanks!

[mrbean-bremen]

Just stumbled over this old question... probably long solved, but I'll comment for anyone else finding this.

Your solution uses the modality LUT to scale the output, which is ok as long it is considered by the viewer (which it should be). If you are only interested in the display in the viewer and not in the actual pixel data, this is sufficient.

The solution from SO was related to the actual pixel data without considering the modality LUT. As you have a typical CT image with an intercept of -1024, your viewer will take that into account and accordingly shift the pixel values, so setting this to 0 is indeed needed if you don't want this. After that, you could either adapt your pixel data as mentioned, or adapt the modality LUT as you did to change the display of the pixel data in the viewer instead.

[CristianoPizzamiglio]

Thanks!