Machine learning for neuroimaging with scikit-learn

Abstract

Statistical machine learning methods are increasingly used for neuroimaging data analysis. Their main virtue is their ability to model high-dimensional datasets, e.g., multivariate analysis of activation images or resting-state time series. Supervised learning is typically used in decoding or encoding settings to relate brain images to behavioral or clinical observations, while unsupervised learning can uncover hidden structures in sets of images (e.g., resting state functional MRI) or find sub-populations in large cohorts. By considering different functional neuroimaging applications, we illustrate how scikit-learn, a Python machine learning library, can be used to perform some key analysis steps. Scikit-learn contains a very large set of statistical learning algorithms, both supervised and unsupervised, and its application to neuroimaging data provides a versatile tool to study the brain.

Keywords: Python; machine learning; neuroimaging; scikit-learn; statistical learning.

Figure 1

Conversion of brain scans into 2-dimensional data.

Figure 2

Maps derived by different methods for face versus house recognition in the Haxby experiment—left: standard analysis; center: SVM weights after screening voxels with an ANOVA; right: Searchlight map. The masks derived from standard analysis in the original paper (Haxby et al., 2001) are displayed in blue and green.

Figure 3

Miyawaki results in both decoding and encoding. Relations between one pixel and four brain voxels is highlighted for both methods. Top: Decoding. Classifier weights for the pixel highlighted [(A) Logistic regression, (C) SVM]. Reconstruction accuracy per pixel [(B) Logistic regression, (D) SVM]. Bottom: Encoding. (E): receptive fields corresponding to voxels with highest scores and its neighbors. (F): reconstruction accuracy depending on pixel position in the stimulus—note that the pixels and voxels highlighted are the same in both decoding and encoding figures and that encoding and decoding roughly match as both approach highlight a relationship between the same pixel and voxels.

Figure 4

Default mode network extracted using different approaches: left: the simple Concat-ICA approach detailed in this article; middle: CanICA, as implemented in nilearn; right: Melodic's concat-ICA. Data have been normalized (set to unit variance) for display purposes.

Figure 5

Brain parcellations extracted by clustering. Colors are random. (A) K-means, 100 clusters, (B) Ward, 100 clusters, (C) K-means, 1000 clusters, and (D) Ward, 1000 clusters.