neuroHarmonize

Harmonization tools for multi-site neuroimaging analysis. Part of the work reported in our paper with data from the ISTAGING consoritum ¹.

If you have identified a bug or are requesting a new feature, please submit an issue on GitHub so that progress can be tracked.

contact (see note above): raymond (dot) pomponio (at) outlook (dot) com

Overview

This package extends the functionality of the package developed by Jean-Philippe Fortin and Nick Cullen ², neuroCombat. New in version 2.4.x: neuroCombat is now a formal dependency.

The reference implementation, neuroCombat, allows the user to perform a harmonization procedure using the ComBat ³ algorithm for correcting multi-site data.

This package, neuroHarmonize, provides similar functionality, with additional features:

Support for working with NIFTI images, with the nibabel package.
Separate train/test datasets.
Specify covariates with generic nonlinear effects. Implemented using Generalized Additive Models (GAMs) from the statsmodels package.
Skip the empirical Bayes (EB) step of ComBat, if desired.

Installation

Latest development version: 2.4.5 (in dev mode)

Requirements:

git >= 2.17.2
python >= 3.6.8

Option 1: Install from PyPI (Stable Version)

pip install neuroHarmonize

Option 2: Install from GitHub (Development Version)

pip install git+https://github.com/rpomponio/neuroHarmonize

Quick Start

Please note: the ComBat ³ algorithm corrects for site effects but intentionally preserves covariate effects. If you wish to remove covariate effects as well you can use the argument return_s_data.

You must provide a data matrix which is a numpy.array containing the features to be harmonized. For example, an array of brain volumes:


  array([[3138.0, 3164.2,  ..., 206.4],
         [1708.4, 2351.2,  ..., 364.0],
         ...,
         [1119.6, 1071.6,  ..., 326.6]])

The dimensionality of this matrix must be: N_samples x N_features

You must also provide a covariate matrix which is a pandas.DataFrame containing all covariates to control for during harmonization. All covariates must be encoded numerically (you must handle categorical covariates in a pre-processing step, see pandas.get_dummies). The DataFrame must also contain a single column called "SITE" with labels that identify sites (the labels in "SITE" need not be numeric).


       SITE   AGE  SEX_M
  0  SITE_A  76.5      1
  1  SITE_B  80.1      1
  2  SITE_A  82.9      0
  ...   ...   ...    ...
  9  SITE_B  82.1      0

The dimensionality of this dataframe must be: N_samples x N_Covariates

The order of samples must be identical in the covariate_matrix and the data_matrix.

The data matrix and the covariate matrix may not contain missing values. If you include missing values in your data such as np.nan, you will encounter errors. Therefore you must perform a complete case analysis. Alterntatively, consider an imputation procedure (beyong the scope of this software, however this implementation may offer a solution).

After preparing both inputs, you can call harmonizationLearn to harmonize the provided dataset.

Example usage:

from neuroHarmonize import harmonizationLearn
import pandas as pd
import numpy as np
# load your data and all numeric covariates
my_data = pd.read_csv('brain_volumes.csv')
my_data = np.array(my_data)
covars = pd.read_csv('subject_info.csv')
# run harmonization and store the adjusted data
my_model, my_data_adj = harmonizationLearn(my_data, covars)

The dimensionality of the matrix my_data_adj will be identical to my_data: N_samples x N_features

Working with NIFTI Images

To work with NIFTI images, you compute a mask using a pandas.DataFrame which contains file paths for all of the images in the training set.

import pandas as pd
import numpy as np
from neuroHarmonize.harmonizationNIFTI import createMaskNIFTI
nifti_list = pd.read_csv('brain_images_paths.csv')
nifti_avg, nifti_mask, affine, hdr0 = createMaskNIFTI(nifti_list, threshold=0)

After the mask is created, you can flatten the images to a 2D numpy.array very similar to what is done with the tabular data example above.

from neuroHarmonize.harmonizationNIFTI import flattenNIFTIs
nifti_array = flattenNIFTIs(nifti_list, 'thresholded_mask.nii.gz')

The next step is identical to working with tabular data. You simply pass the 2D array to neuroHarmonize.harmonizationLearn.

import neuroHarmonize as nh
covars = pd.read_csv('subject_info.csv')
my_model, nifti_array_adj = nh.harmonizationLearn(nifti_array, covars)
nh.saveHarmonizationModel(my_model, 'MY_MODEL')

Lastly, you can apply the model sequentially to images in a larger dataset with applyModelNIFTIs. When performing NIFTI harmonization, loading the entire set of images may exceed memory capacity. This function will reduce the burden on memory by applying the model to images one-by-one and saving the results as NIFTIs.

from neuroHarmonize.harmonizationNIFTI import applyModelNIFTIs
# load pre-trained model
my_model = nh.loadHarmonizationModel('MY_MODEL')
applyModelNIFTIs(covars, my_model, nifti_list, 'thresholded_mask.nii.gz')

Applying Pre-Trained Models to New Data

This feature allows you to train a harmonization model on a subset of data, then apply the model to the entire set. For example, in longitudinal analyses, one may wish to train a harmonization model on baseline cases and apply the model to follow-up cases, to avoid double-counting subjects.

If you have previously trained a harmonization model with harmonizationLearn, you may apply the model parameters to new data with harmonizationApply.

First load the model:

from neuroHarmonize import harmonizationApply, loadHarmonizationModel
import pandas as pd
import numpy as np
# load a pre-trained model
my_model = loadHarmonizationModel('../models/my_model')

Next, prepare the holdout data on which you will apply the model. This data must look exactly like the training data for harmonizationLearn, including the same number and order of covariates. If the holdout data contains a different number of sites, an error will be thrown.

After preparing the holdout data simply apply the model:

df_holdout = pd.read_csv('../data/brain_volumes_holdout.csv')
my_holdout_data = np.array(df_holdout)
covars = pd.read_csv('subject_info_holdout.csv')
my_holdout_data_adj = harmonizationApply(my_holdout_data, covars, my_model)

Specifying Nonlinear Covariate Effects

You may specify nonlinear covariate effects with the optional argument: smooth_terms. For example, you may want to specify age as a nonlinear term in the harmonization model, if age exhibits nonlinear relationships with brain volumes. This can be done easily with harmonizationLearn:

from neuroHarmonize import harmonizationLearn
import pandas as pd
import numpy as np
# load your data and all numeric covariates
my_data = pd.read_csv('brain_volumes.csv')
my_data = np.array(my_data)
covars = pd.read_csv('subject_info.csv')
# run harmonization with NONLINEAR effects of age
my_model, my_data_adj = harmonizationLearn(data, covars, smooth_terms=['AGE'])

When applying nonlinear models to holdout data, you may get an error: "some data points fall outside the outermost knots, and I'm not sure how to handle them". This is documented: statsmodels/statsmodels#2361. The current workaround is to use the optional argument: smooth_term_bounds, which controls the boundary knots for nonlinear estimation. You should specify boundaries that contain the limits of the entire dataset, including holdout data.

Empirical Bayes

Note the default behavior is to run the empirical Bayes (EB) step of ComBat, which is useful for harmonizing multiple features that are similar such as genes or brain regional volumes.

To run without EB, simply pass the optional argument eb=False to harmonizationLearn. This is convenient when harmonizing a small number of features, e.g. fewer than 10.

When eb=True, ComBat uses Empirical Bayes to fit a prior distribution for the site effects for each site. You may wish to visualize fit of the prior distribution, along with the observed distribution of site effects. The following code example plots both distributions for the location effect of site 1.

import matplotlib.pyplot as plt
import seaborn as sns
from neuroHarmonize import loadHarmonizationModel
model = loadHarmonizationModel('../models/my_model')
site_01 = stats.norm.rvs(size=10000, loc=model['gamma_bar'][0], scale=np.sqrt(model['t2'][0]))
sns.kdeplot(site_01, color='blue', label='Site-1-prior')
sns.kdeplot(model['gamma_hat'][0, :], color='blue', label='Site-1-observed', linestyle='--')
plt.show()

Reference Sites

Recently other implementations have been including the ability to specify a reference site or scanner, such that all data will be harmonized to that site.

This feature was added in version 2.4.x. It can be used as in the following example:

from neuroHarmonize import harmonizationLearn
import pandas as pd
import numpy as np
# load your data and all numeric covariates
my_data = pd.read_csv('brain_volumes.csv')
my_data = np.array(my_data)
covars = pd.read_csv('subject_info.csv')
# run harmonization and store the adjusted data
my_model, my_data_adj = harmonizationLearn(my_data, covars, ref_batch='SITE_A')

Please submit any issues to GitHub if you run into trouble with this feature, as I am still testing its utility.

Citations

Pomponio, R., Shou, H., Davatzikos, C., et al., (2019). "Harmonization of large MRI datasets for the analysis of brain imaging patterns throughout the lifespan." Neuroimage 208. https://doi.org/10.1016/j.neuroimage.2019.116450. ↩
Fortin, J. P., N. Cullen, Y. I. Sheline, W. D. Taylor, I. Aselcioglu, P. A. Cook, P. Adams, C. Cooper, M. Fava, P. J. McGrath, M. McInnis, M. L. Phillips, M. H. Trivedi, M. M. Weissman and R. T. Shinohara (2017). "Harmonization of cortical thickness measurements across scanners and sites." Neuroimage 167: 104-120. https://doi.org/10.1016/j.neuroimage.2017.11.024. ↩
W. Evan Johnson and Cheng Li, Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics, 8(1):118-127, 2007. https://doi.org/10.1093/biostatistics/kxj037. ↩ ↩²

rpomponio/neuroHarmonize