Nitrain-image

Nitrain-image is a library that lets you read / write, visualize, and operate on medical imaging data in a natural, pythonic way. If you've used numpy before, then nitrain-image will be easy to learn. All of the most common image formats - dicom, nifti, etc. - are supported.

If you are interested in training medical imaging AI models, then you may want to check out the nitrain library. You can also learn more about using nitrain-image and nitrain in the open-source book "Becoming a medical imaging AI expert with Python".

Quickstart

Here is a 10 minute introduction to the nitrain-image library. If you have used numpy before, that is about all you will need to start working with medical images like a pro.

To begin, the easiest way to install the nitrain-image package is from PyPi using pip.

Installation

pip install nitrain-image

Loading and saving

You can load basically any type of medical image using nitrain-image.

import ntimage as nti
img = nti.load('image.nii.gz')

Saving is equally easy.

nti.save(img, 'image.nii.gz')

Creation

If you have a numpy array without any medical imaging metadata, you can create an ntimage.

import numpy as np
img = nti.from_numpy(np.random.randn((128,128)))

Converting any ntimage - even those read from file - to a numpy array can also be done.

arr = img.numpy()

If you have an ntimage with associated metadata such as origin, spacing, direction, and so on, then it is possible to transfer that info when creating an ntimage from a numpy array.

img.set_origin((10,10))
img2 = nti.from_numpy_like(np.random.randn((128,128)), img)

As with numpy or torch, you can also create an ntimage in a variety of convenient ways:

img = nti.ones((128,128))
img = nti.zeros((128,128))
img = nti.rand((128,128))

The ntimage class has the same datatypes as numpy arrays. If you want to cast an ntimage to another datatype, you can do so with the astype function.

img = img.astype('float32')

Indexing

You can index an ntimage as you would a numpy array, but keep in mind that indexing an ntimage returns another ntimage and NOT a numpy array. This makes it very convenient to crop images.

img = nti.ones((128,128))
img[:20,:20]

Assignment of values is also possible via indexing.

img = nti.ones((128,128))
img[:20,:20] = 0

Indexing an ntimage with another ntimage is also supported. In this case, the image you use to index will be treated like a mask and the resulting image will have any values outside of the index set to zero.

img = nti.rand((128,128))
mask = nti.zeros((128,128))
mask[:20,:20] = 1
img2 = img[mask] # all indices outside of (:20, :20) will now be zero

Math operations

As with numpy, you can use any core math operation on ntimages.

img = nti.ones((128,128))
img2 = nti.ones((128,128))

img = img * 10
img = img + img2
img = img / 10

There are also many other math functions such as exp and log that can be applied to an ntimage. Additionally, all summary functions such as median, mean, sum, min, max, and so on, are available.

img = nti.exp(img)
img = nti.log(img)

Logical functions will also work as expected on ntimages.

img = nti.ones((128,128))
img2 = nti.rand((128,128)) > 0.5

img3 = img & img2

Image operations

You can use many traditional image procesing operations on ntimages. The smooth function is one example.

img = nti.load(nti.example_data('r16'))
img2 = nti.smooth(img, sigma=2)

The resample function also has many important use cases for ntimages.

img2 = nti.resample(img, (64,64))

Because nitrain-image is built on top of ITK, you can build you own custom image processing function using the itk package in Python and integrate it using the from_itk function.

img = nti.load(nti.example_data('r16'))

def my_function(image):
    itk_image = image._image
    # perform some itk processing...
    # ...
    return from_itk(itk_image)

The interoperability of nitrain-image with numpy and itk make it easy to do pretty much anything.

Plotting

The visualization functions of nitrain-image are highly intuitive with sensible defaults, but they are still flexible. The plot function will take care of most needs.

img = nti.load(nti.example_data('mni'))
nti.plot(img)

The plot function also supports overlays. This is useful for showing labels on top of your images.

img2 = img > img.median()
nti.plot(img, img2)

Other functions such as grid and plot_ortho provide you with more flexibility to create unique, publication-quality visualizations from your images.

Lastly, the plot_hist function can come in handy to view the distribution of an image's intensity values.

img = nti.load(nti.example_data('r16'))
nti.plot_hist(img)

Contributing

If you have a bug to report or are interested in contributing new features to nitrain-image, feel free to create an issue here on GitHub. The nitrain-image library builds upon the excellent Insight Toolkit (ITK).