Convert NIfTI images into multiple TIF images

nii_to_tif.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
nii_to_tif

command line executable to convert 3d nifti images to 
individual tiff images along a user-specified axis

call as: python nii_to_tif.py /path/to/nifti /path/to/tif
(append optional arguments to the call as desired)

Author: Jacob Reinhold (jacob.reinhold@jhu.edu)
"""

import argparse
from glob import glob
import os
import sys

from PIL import Image
import nibabel as nib
import numpy as np


def arg_parser():
    parser = argparse.ArgumentParser(description='split 3d image into multiple 2d images')
    parser.add_argument('img_dir', type=str, 
                        help='path to nifti image directory')
    parser.add_argument('out_dir', type=str, 
                        help='path to output the corresponding tif image slices')
    parser.add_argument('-a', '--axis', type=int, default=2, 
                        help='axis of the 3d image array on which to sample the slices')
    parser.add_argument('-p', '--pct-range', nargs=2, type=float, default=(0.2,0.8),
                        help=('range of indices, as a percentage, from which to sample ' 
                              'in each 3d image volume. used to avoid creating blank tif '
                              'images if there is substantial empty space along the ends '
                              'of the chosen axis'))
    return parser


def split_filename(filepath):
    path = os.path.dirname(filepath)
    filename = os.path.basename(filepath)
    base, ext = os.path.splitext(filename)
    if ext == '.gz':
        base, ext2 = os.path.splitext(base)
        ext = ext2 + ext
    return path, base, ext


def main():
    try:
        args = arg_parser().parse_args()
        fns = glob(os.path.join(args.img_dir, '*.nii*'))
        for fn in fns:
            _, base, ext = split_filename(fn)
            img = nib.load(fn).get_data().astype(np.float32).squeeze()
            if img.ndim != 3:
                print(f'Only 3D data supported. File {base}{ext} has dimension {img.ndim}. Skipping.')
                continue
            start = int(args.pct_range[0] * img.shape[args.axis])
            end = int(args.pct_range[1] * img.shape[args.axis])
            for i in range(start, end):
                I = Image.fromarray(img[i,:,:], mode='F') if args.axis == 0 else \
                    Image.fromarray(img[:,i,:], mode='F') if args.axis == 1 else \
                    Image.fromarray(img[:,:,i], mode='F')
                I.save(os.path.join(args.out_dir, f'{base}_{i:04}.tif'))
        return 0
    except Exception as e:
        print(e)
        return 1


if __name__ == "__main__":
    sys.exit(main())

@mccrinbc you can use all three orientations as input to the network or train three separate models; there are papers that have successfully done both, and I'm not sure which is more optimal—the optimal choice would depend on the task/resource-availability. Let me know if my response didn't answer your question.

Thanks for the reply Jacob!

It makes sense that you could use either method depending on the problem you're trying to solve. But, does it make sense to classify all of these slices from a single brain as one class? In the case of a binary classification problem, depending on the disease pathology, there could be sections of the brain that are unaffected. From this, you could have a large amount of data "incorrectly classified". Have you classified all of the slices you've obtained from nii_to_tif.py as a single class, or have you been more meticulous to further label your data based on what's shown within the slice? Or is this interpretation flawed?

This is very specific, so please feel free to remove this discussion as you see fit. Thank you!

@mccrinbc I guess I'm confused as to what you mean by classifying "all slices from a single brain as one class." I think you have a specific application in mind, e.g., classifying an image as diseased or not diseased; however, there are many reasons to create 2D slices of a 3D image, and this script is not application-specific. For example, I've used this script to create 2D slices for input to a deep neural network for segmentation and image-to-image translation. If you have registered labels and you use this script to create slices, using the same options, with the source and label/target image, you'll maintain the correspondence between voxels/pixels.

If you only have a weak label like: "this image contains disease," I think your intuition is correct that it would be flawed to use this label on all slices because some of the slices may only contain healthy-appearing tissue. This particular problem isn't my research area, but I'm familiar with some work dealing with this problem, e.g., this paper.

Hope that helps.

Jacob,
This was incredibly helpful, thank you for your insight!

I was wondering something. I thought that this was an amazing tool overall, however, it appears that my tif files aren't coming out too well. However, when I convert the tif files back to nii (with the other code), it comes out perfect and normal. Do you know what might be the problem?

^^Here is what I got

^^Here is what I was expecting

THANK YOU so much for all your work! It really helps me with my research!

@saranggoel Can you upload the output TIF image? Perhaps your window-level setting is bad. Or have you already looked at that?

Given that the nifti image reconstructed from the TIF images is fine, I'm almost sure that this problem is just a window-level setting problem. Open up the TIF image with some problem like mipav, some other medical imaging viewer, or matlab/matplotlib and adjust the window-level settings and you should see the image you're expecting.

I'm guessing you are trying to view the image with a GUI-based image viewer in whatever your operating system is. However, the default window-level settings probably don't work for the TIF image which has a large range of intensitites (i.e., hounsfield units).

You shouldn't use this script if you want to convert to jpg as the final output though. The reason I built this script is that I wanted to convert a nifti image into a collection of 2d images while preserving the exact intensity value at every voxel. When you convert to jpg or png you need to quantize the intensity values (i.e., convert them to uint8) which is—in general—a lossy process.

Hope this helps.

Thank you for your prompt reply! Yes, opening it in a medical viewer software made it look exactly as expected! Thank you so much!

Hi Jacob!
I am trying to run the code but I keep encountering the following error:

$ python nii_to_tif.py brainmask/ TIF/
File "nii_to_tif.py", line 125
print(f'Only 3D data supported. File {base}{ext} has dimension {img.ndim}. Skipping.')
^
SyntaxError: invalid syntax

Do you know what might be the error? I have tried different 3D nifti files but none of them work.
brainmask/ is a folder containing a 3D nifti file
TIF/ is an empty folder
Thank you very much!

Actually, there is some syntax error with your print statement in the code. If you are very sure your nii files are 3D, then just delete the following lines:

        if img.ndim != 3:
            print(f'Only 3D data supported. File {base}{ext} has dimension {img.ndim}. Skipping.')
            continue

Hi Jacob!
I am trying to run the code but I keep encountering the following error:

$ python nii_to_tif.py brainmask/ TIF/
File "nii_to_tif.py", line 125
print(f'Only 3D data supported. File {base}{ext} has dimension {img.ndim}. Skipping.')
^
SyntaxError: invalid syntax

Do you know what might be the error? I have tried different 3D nifti files but none of them work.
brainmask/ is a folder containing a 3D nifti file
TIF/ is an empty folder
Thank you very much!

@yastrip The reason you are receiving a syntax error is because you are using a version of python <3.6. The f-string format was added in python 3.6, so previous versions will not recognize the syntax. You can either remove the like as suggested by @saranggoel, or you can change the print statement to:

print('Only 3D data supported. File {}{} has dimension {}. Skipping.'.format(base, ext, img.ndim))

A similar question was already asked. See here. Hope this helps.

Hi, I'm trying to convert 3 dim. nii file to multiple images with your script but it only gives multiple blank tiff files. When I convert nii to png files everything is fine. Can you help me about that?

Hi @ozturkoktay, the TIFF images are probably fine. You just need to open them in a medical image viewer like MIPAV or plot them in matplotlib. This has been discussed above here and here.

Hey Jacob,
I would like to confirm that the conversion from nii to tiff is lossless. Is there a simple way to do this?

Thanks!
Brian

@mccrinbc Load your nifti image with nibabel and extract the data (.get_fdata(dtype=np.float32)). Open up all the corresponding tiff images and concatenate them along the appropriate axis. Check for equality with np.allclose(nifti_data, tiff_data))

Note that this conversion is only lossless (up to numerical precision) if the original nifti image is 32-bit floating point.

Beauty - Thanks dude!

Hello! This script is very helpful. The code worked well.
However, I have a question.
My nifti file is z-normalized. So its pixel range has negative values and doesn't go up to 255 like other image arrays.
As my purpose is to use imagenet transfer learning model (such as resnet, inception etc), I have to make input shape to 3 channel (ex; (220,220,3) ). So I did so. However when I show this 3 channel image through matoplotlib, it looks bad. If I show each channel alone, it shows fine.
Is z-normalized the problem? Thank you.

-3 channel image

-only one channel shown (image[:,:,1])

@sujeongEOM see up above (e.g., here) for a discussion of this issue.