MAIA Tools

A collection of image processing tools designed for the Erasmus Mundus Joint Masters Degree in Medical Imaging and Applications (MAIA)

Content

• Introduction
• Requirements
• Classes
• Examples
• License

Introduction

This repository contains tools for image registration and segmentation implemented and used for my university projects. The tools were extensively used on Brain MRI Images as well as lung ct scans. Focus was to provide easy to use interfaces with common imaging libraries as well as implementation of custom algorithms for clustering and (Multi)Atlas segmentation.

The core of the project is the ImageInterface Object. The idea is to make it possible to load images with whatever library the user wants and then internally convert it to the necessary types used. This is inspired by my own experience when the sitk SimpleElastix library failed to build locally, leading to rebasing the entire project to itkElastix.

Requirements

This package can be installed from PyPI by running the following command in your terminal:

Classes

A description of the classes in the package

GMM

The GaussianMixtureModel class, supports kmeans, atlas, tissue model and combined initialization. It is a clustering algorithm that fits n clusters to a data distribution. The clusters are Gaussian distributions, which allows this algorithm to model elliptical data distributions. Optimizes the clusters using the Expectation Maximization algorithm Support posterior assignment using prior info from Atlas or Tissue Model Supports informed EM by applying prior information on the weights after the E-Step. Example usage in the example script.

Parameters:

• k: int, number of clusters
• prior = Prior knowledge object. Must support a segment(image) and a soft_segment(image) function, optional
• max_iter int, number of iterations to stop at if convergence is not reached, default = 100
• verbose, bool, controls if detailed parameter changes are printed to the console, default False
• init, String, the initialization strategy, supports 'kmeans' or 'prior' init. Note that usage of posterior and influence requires prior init for label consistency. Optional, default = 'kmeans'

Methods:

fit(image, mask, influence_frq)

Fits the gaussian distributions to the dataset.

• image: 3D numpy array, the image to be segmented, can be a list of images to combine multiple modalities
• mask: 3D numpy array, binary image to provide an area of interest to perform skullstripping
• influence_frq: int, the frequency of prior knowledge influence on the expectation maximization. Frequency in iterations. Optional, default = 0 -> no influence

fit_transform(image, mask, influence_frq, posteriori) -> segmented image

Same as fit but returns the segmented image after it is done.

• posteriori: bool, use posterior influence or not, default = False

Scorer

Brute force label reassignment. Used in case of kmeans initialization to fit the randomly ordered labels to the gt. Computes the dice score for every possible label combination and picks the best by finding the largest average dice score across labels

Parameters:

• gt: 3D numpy array, the labeled ground truth
• predictions: 1xN array of assigned samples, output of the em, not really used anymore. Optional
• image: 3D numpy array, the labeled output of the segmentation algorithm

Methods

relabel() -> image relabeled, _, _

performs the scoring and returns the best result return:

• img: 3D numpy array, the relabeled segmentation image
• predictions: 1xN array, the relabeled predictions, can be discarded
• score: list of float, the dice scores for each label in the best combination

FeatureTransformer

Transforms the data from image into feature space. Used internally in the GMM class.

Parameters:

None

Methods:

transform(images, mask) -> list of data

Transforms the volumetric data into the feature space for the GMM. Uses an area of interest to generate the data, essentially performing the skull stripping

• images: list of 3D numpy arrays, the sources of the data, can be multi modality or 3D probability maps for atlas
• mask: 3D bool numpy array, mask for the area of interest

retransform(images) -> list of data

Transforms volumetric data into feature space by using the internally stored area of interest, passed from the transform method ensures consistency in data order.

• images: list of 3D numpy arrays, the sources of the data, can be multi modality or 3D probability maps for atlas

reverse(labels) -> volumetric image

Transforms data from feature space back into 3D space. Takes a list of labels and moves them to the corresponding voxel.

• labels: 1xN numpy array, labeled datapoints in the same order as the transformed output

Atlas

The atlas object, builds or loads a probabilistic atlas from a set of registered images and ground truths. Models the probability of a voxel to be a certain tissue by using its spatial location.

Parameters

• labels: int, the number of labels in the ground truth, optional, default = 4 for bg, csf, wm, gm

Methods

build(images, gts, affine)

Build a probabilistic and anatomical atlas from an ordered list of registered images and ground truths as well as a reference image.

• images: list of 3D numpy arrays, the registered images of the atlas training set
• gts: list of 3D numpy arrays, the registered ground truths of the atlas training set
• affine: 2D numpy array, the affine transformation matrix of the fixed image of the atlas registration.

save(path)

Makes a directory and stores the atlas as nifty files.

• path: String, the path to the save location

load(path)

Loads an atlas from disk to use in code.

• path: String, the path to the save location

segment(image) -> labeled image

Returns the anatomical atlas. If the atlas is registered to a target image, this corresponds to the segmentation of the image.

• image: discarded parameter for compatibility with the other prior knowledge objects.

soft_segment(image) -> probability image

Retruns the probabilistic atlas for the tissues csf, wm and gm. If the atlas is registered to target image, this corresponds to the tissue probabilities for the image.

• image: discarded parameter for compatibility with the other prior knowledge objects.

TissueModel

The tissue model object, builds or loads a tissue model from a set of unregistered images and ground truths. Models the probability of a voxel to be a certain tissue by using its intensity.

Parameters

• labels: int, the number of tissues in the ground truth, optional, default = 3 for csf, gm, wm
• norm: String, the normalisation strategy, optional, default='linear', supports 'linear' for linear histogram equalization, 'hist' for histogram matching, or 'none' for no equalization.

Methods

fit(images, gts)

fits the tissue model to the data. The images and labels do not need to be registered, but the order of images and gts must be the same.

• images: list of 3D numpy arrays, the images used in the dataset
• gts: list of 3D numpy arrays, the labels used in the dataset

save(path)

Makes a directory and stores the tissue model as numpy files.

• path: String, the path to the save location

load(path)

Loads a tissue model from a save location

• path: String, the path to the save location

segment(image) -> labeled image

Segment an image by assigning each voxel a label based on its intensity and the tissue model.

• image: 3D numpy array, the image to segment

soft_segment(image) -> probability image

Generate a probability map by assigning each voxel a tissue probability for each tissue based on its intensity.

• image: 3D numpy array, the image to segment

show()

Makes a plot of the tissue model, intensity vs probability for each label

Examples

A set of example scripts that show the usage of the provided code

Atlas building

Shows how to build an atlas using the provided class. Note that the images need to be registered beforehand. Usage of the provided parameter files is recommended.

Tissue Model building

Shows how to build a tissue model using the provided class. Does not require any registration.

Extended GMM usage

Shows how to segment an image using the provided class. Note that the atlas has to be registered to each target image individually beforehand.

License

This project is licensed under the MIT License - see the LICENSE file for details.