How do I convert Pixel Data to a PNG image?

[yehudahkay]

The attached file is converted to a black image rather than the actual image using the following code:

ds = dcmread(filepath)
im = fromarray(ds.pixel_array)
im.save(folder + filename + '.png')

https://drive.google.com/file/d/1DwJNQDChz3J79ylqgpSG7RUwvlhAIG5P/view?usp=sharing

[mrbean-bremen]

From a quick look at the dataset I would guess that you should apply the modality LUT and the VOI LUT before converting it to a PNG.

[yehudahkay]

I tried the following code but got the same results, do i need to do something differently? Thanks!

ds = dcmread(filepath)
pydicom.pixels.processing.apply_modality_lut(ds.pixel_array, ds)
pydicom.pixels.apply_voi_lut(ds.pixel_array, ds)
#save jpg
im = fromarray(ds.pixel_array)
im.save(folder + filename + '.png')

[scaramallion]

Every time you're calling Dataset.pixel_array a new ndarray of the original Pixel Data is being created, so don't try to modify it in-place.

ds = dcmread(filepath)
arr = ds.pixel_array
arr = pydicom.pixels.processing.apply_modality_lut(arr, ds)
arr = pydicom.pixels.apply_voi_lut(arr, ds)
#save jpg
im = fromarray(arr)
im.save(folder + filename + '.png')

[yehudahkay]

I tried this but I still only get a completely black png. This file can be opened with a dcm viewer and it displays a medical image.

[scaramallion]

(0002,0002) UI =MRImageStorage                          #  26, 1 MediaStorageSOPClassUID
(0002,0010) UI =LittleEndianExplicit                    #  20, 1 TransferSyntaxUID
...
(0028,0002) US 1                                        #   2, 1 SamplesPerPixel
(0028,0004) CS [MONOCHROME2]                            #  12, 1 PhotometricInterpretation
(0028,0010) US 512                                      #   2, 1 Rows
(0028,0011) US 512                                      #   2, 1 Columns
(0028,0030) DS [0.44921875\0.44921875]                  #  22, 2 PixelSpacing
(0028,0100) US 16                                       #   2, 1 BitsAllocated
(0028,0101) US 12                                       #   2, 1 BitsStored
(0028,0102) US 11                                       #   2, 1 HighBit
(0028,0103) US 0                                        #   2, 1 PixelRepresentation
(0028,0301) CS [NO]                                     #   2, 1 BurnedInAnnotation
(0028,0303) SH [MODIFIED]                               #   8, 1 LongitudinalTemporalInformationModified
(0028,1050) DS [1098]                                   #   4, 1 WindowCenter
(0028,1051) DS [1909]                                   #   4, 1 WindowWidth
(0028,1052) DS [0]                                      #   2, 1 RescaleIntercept
(0028,1053) DS [5.92258852258852]                       #  16, 1 RescaleSlope
(0028,1054) LO [normalized]                             #  10, 1 RescaleType
(0028,2110) CS [00]                                     #   2, 1 LossyImageCompression

Could you post a reproducible example, I don't know what fromarray() is from, (maybe pillow?)

[yehudahkay]

import pydicom
from pydicom import dcmread
from pydicom.pixel_data_handlers.util import apply_modality_lut
from PIL.Image import fromarray

ds = dcmread(filepath)
arr = ds.pixel_array
pydicom.pixels.apply_voi_lut(arr, ds)
pydicom.pixels.processing.apply_modality_lut(arr, ds)
#save jpg
im = fromarray(arr)
im.save(folder + filename + '.png')

[scaramallion]

You can't save a float array as PNG, it needs to be converted to ints. Also, that isn't the example I posted, those functions return a new ndarray instance (as it says in the documentation).

The other probable issue is you're just blinding converting some data to PNG without rescaling it. If you're using 16-bit PNGs then your actual image data range will be ~[0, 3542], so of course it will appear dark.

[yehudahkay]

@scaramallion thank you so much. I tried with the correct code but i get the error OSError: cannot write mode F as JPEG

[scaramallion]

You can't write floats to JPEG either. Pretty much all image formats require ints in a specific range.

[scaramallion]

Simple example for 16-bit PNGs:

from pydicom import dcmread
from PIL import Image

ds = dcmread("2122.dcm")
arr = ds.pixel_array
# Convert to floats
f_arr = arr.astype(float)
 # Rescale to use full 16-bit range
f_arr /= arr.max()
f_arr *= (2**16 - 1)
print(f_arr.max())  # 65535.0
# Convert to uint16
u16_arr = f_arr.astype("u2")
print(u16_arr.max())  # 65535

im = Image.fromarray(u16_arr)
im.save("2122.png")

[yehudahkay]

@scaramallion this is amazing but how do I know if this type of conversion is necessary or not from the ds data? I'm looking through but there are many properties

[scaramallion]

If you're converting Pixel Data to any sort of imaging format like PNG or JPEG then some form of manipulation of the raw pixel data will always be necessary.

[yehudahkay]

Ok so should I be checking if bits allocated is 8 or 16? Are there other options besides these which require another conversion? Thanks!

[scaramallion]

You may also need to account for Pixel Representation, which tells you whether the Pixel Data uses signed or unsigned integers (although you could just check the ndarray's dtype instead). Depending on your use case you may also need to apply windowing.

Bits Stored will tell you the theoretical Pixel Data value range, but again you could just check the max/min values in the ndarray.