Back

Machine learning for neuroimaging with scikit-learn

Alexandre AbrahamFabian PedregosaMichael EickenbergPhilippe GervaisAndreas MullerJean KossaifiAlexandre GramfortBertrand ThirionGäel Varoquaux

article2014Front. Neuroinform.2,197 citations

Demonstrates how to apply scikit-learn to functional neuroimaging datasets, providing practical implementations for high-dimensional brain decoding, encoding, and resting-state fMRI analysis.

Listen

Functional brain imaging generates massive, high-dimensional datasets where the number of measured brain locations vastly exceeds the number of observations. While statistical machine learning provides the necessary analytical power to model these complex relationships, adopting these techniques is often hindered by steep technical barriers and a divide between specialized neuroscience questions and general computer science software. Bridging this gap requires accessible, standardized computational frameworks that allow researchers to extract meaningful neural patterns without relying on inflexible black-box systems.

The article aims to demonstrate how a general-purpose Python machine learning toolkit, scikit-learn, can perform core functional neuroimaging analyses through clean, modular, and interpretable workflows.

The authors evaluate this framework across supervised and unsupervised functional magnetic resonance imaging applications using public neuroimaging benchmarks. Their approach details practical workflows for transforming complex four-dimensional brain scans into standardized two-dimensional matrices, applying signal cleaning steps such as detrending, and running learning algorithms. They examine decoding brain states, encoding stimulus features into neural activations, and mapping functional connectivity networks during resting-state conditions.

The findings show that standard linear classifiers combined with simple univariate feature screening successfully isolate discriminative visual cortex regions, matching classical findings while providing predictive capabilities for unseen scans. In visual reconstruction tasks, sparse linear models achieve cross-validation accuracy rates of approximately 70% under optimal regularization settings, accurately mapping receptive fields in primary visual areas with higher reconstruction fidelity near the fovea. In unsupervised resting-state analysis, independent component analysis and agglomerative hierarchical clustering successfully uncover coherent large-scale networks, such as the default mode network, as well as spatially contiguous functional parcels.

These results demonstrate that neuroimaging workflows can achieve high predictive performance and clear neurobiological interpretability using standard, versatile open-source tools rather than complex custom code. This integration streamlines analysis pipelines, reduces technical overhead, improves research reproducibility, and supports biomarker discovery in clinical cohorts where task-based scans are not feasible.

To fully leverage these methods, research teams should adopt standardized scientific Python pipelines, incorporate rigorous cross-validation for hyperparameter selection, and utilize domain-specific wrappers that streamline spatial masking and template registration. While general-purpose tools are highly effective, users must exercise caution regarding data preprocessing artifacts, the loss of spatial context during matrix flattening, and the computational cost of spatial searchlight procedures, ensuring robust validation across independent cohorts.

CEAImperial College LondonINRIATélécom ParisUniversity of Bonnlibrariespythonstatistical learningneuroimaging machine learningfunctional magnetic resonance imagingfMRI time-series
arXiv: 1412.3919AlexandreAbraham/frontiers2013
  • Paper: Scikit-learn: Machine Learning in Python, Fabian Pedregosa et al. (2011). Reading this foundational documentation on scikit-learn's API design and core modules provides the essential software library background required to implement the machine learning neuroimaging workflows discussed in the source.
  • Paper: Decision-Making with Auto-Encoding Variational Bayes, Romain Lopez et al. (2020). This ebook extends the statistical modeling foundations established in the source by examining advanced variational auto-encoding Bayes methods for downstream decision-making tasks.
Cover for 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.

Topics

librariespythonstatistical learningsupervised machine learningunsupervised learninghigh resolution imagecorrelation analysis

Keywords

neuroimaging machine learningfunctional magnetic resonance imagingfMRI time-seriesbrain mappingsupport vector machinesindependent component analysisstatistical learningfeature selectionspectral clustering

Table of Contents

  • 1
  • 2 Keywords:
  • 3 Introduction
  • 4 Our tools: scikit-learn and the Python ecosystem
  • 4.1 Basic scientific Python tools for the neuroimager
  • 4.2 Scikit-learn and the machine learning ecosystem
  • 4.3 Scikit-learn concepts
  • 5 Data preparation: from MR volumes to a data matrix
  • 5.1 Spatial resampling
  • 5.2 Signal cleaning
  • 5.3 From 4-dimensional images to 2-dimensional array: masking
  • 5.4 Data visualisation
  • 6 Decoding the mental representation of objects in the brain
  • 6.1 Classification with feature selection and linear SVM
  • 6.2 Searchlight
  • 6.3 Results
  • 7 Encoding brain activity and decoding images
  • 7.1 Decoding
  • 7.2 Encoding
  • 7.2.1 Receptive fields
  • 7.3 Results
  • 8 Resting-state and functional Connectivity analysis
  • 8.1 Independent Component Analysis (ICA) to extract networks
  • 8.1.1 ICA in neuroimaging
  • 8.1.2 Application
  • 8.1.3 Results
  • 8.2 Learning functionally homogeneous regions with clustering
  • 8.2.1 Approaches
  • 8.2.2 Results
  • 9 Conclusion
  • References

Knowls

  1. Knowl 1 — Supervised fMRI Decoding via Univariate Feature Screening and Linear Support Vector Machines

    model/method

    In functional neuroimaging decoding, a machine learning model is trained to predict behavioral states, stimulus categories, or clinical variables yy from high-dimensional fMRI activation scans X∈Rn×pX \in \mathbb{R}^{n \times p}, where the number of features (voxels, e.g., p≈40,000p \approx 40{,}000) vastly exceeds the number of samples (time volumes or trials, e.g., n≈1,400n \approx 1{,}400).

    To overcome the curse of dimensionality and extract interpretable neural representations, a two-stage pipeline is used:

    1. Univariate Feature Selection: An ANOVA FF-test evaluates the null hypothesis that each voxel's activation level is independent of the target class yy. A fixed number of top discriminative voxels (e.g., k=500k = 500) or a percentile threshold is selected using a feature selection transformer.

    2. Linear Classification: A Support Vector Classifier (SVC) with a linear kernel finds the optimal separating hyperplane in the reduced feature space by solving: min⁡w,b12∥w∥22+C∑i=1nmax⁡(0,1−yi(w⊤xi+b))\min_{w, b} \frac{1}{2} \|w\|_2^2 + C \sum_{i=1}^n \max(0, 1 - y_i (w^\top x_i + b)) where w∈Rkw \in \mathbb{R}^k is the weight vector, b∈Rb \in \mathbb{R} is the intercept, and C>0C > 0 is the regularization parameter.

    3. Spatial Weight Interpretation: Because a linear kernel defines the decision boundary directly in feature space, the learned coefficients ww can be projected back into the original 3D brain volume using inverse masking. The resulting discriminative weight map identifies brain regions (such as house- or face-selective visual areas in ventral temporal cortex) that drive prediction.

  2. Knowl 2 — Cross-Validated Predictive r-squared Metric for Encoding Models

    equation

    In functional neuroimaging encoding models, where fMRI BOLD signals across brain voxels are predicted from external stimulus representations, prediction quality on held-out test data is quantified using the predictive r2r^2 score (coefficient of determination):

    r2=1−∑i=1ntest(yi−y^i)2∑i=1ntest(yi−yˉtest)2r^2 = 1 - \frac{\sum_{i=1}^{n_{\text{test}}} (y_i - \hat{y}_i)^2}{\sum_{i=1}^{n_{\text{test}}} (y_i - \bar{y}_{\text{test}})^2}

    where yi∈Ry_i \in \mathbb{R} is the true observed BOLD signal of a given voxel for test sample ii, y^i∈R\hat{y}_i \in \mathbb{R} is the corresponding out-of-sample predicted BOLD response from the fitted encoding model (such as Ridge or Lasso regression), ntestn_{\text{test}} is the number of samples in the test fold, and yˉtest=1ntest∑i=1ntestyi\bar{y}_{\text{test}} = \frac{1}{n_{\text{test}}} \sum_{i=1}^{n_{\text{test}}} y_i is the empirical mean of the observed test responses.

    A score of r2=1.0r^2 = 1.0 corresponds to perfect prediction; r2=0.0r^2 = 0.0 indicates that the model performs no better than predicting the mean of the test set; and r2<0.0r^2 < 0.0 indicates that out-of-sample model predictions perform worse than the constant baseline mean. Computing this score independently across voxels identifies the retinotopic or sensory regions whose variance is captured by the stimulus model.

  3. Knowl 3 — Population Receptive Field Estimation via Sparse Linear Regression

    model/method

    To estimate localized population receptive fields (pRF) in early visual cortex, the activation xv∈Rnx_v \in \mathbb{R}^n of a single brain voxel vv across nn visual stimulus presentations is modeled as a linear combination of stimulus pixel intensities Y∈Rn×mY \in \mathbb{R}^{n \times m} (e.g., m=10×10=100m = 10 \times 10 = 100 binary pixels).

    Because retinotopy dictates that individual voxels respond selectively to a small, contiguous portion of the visual field, the weights β∈Rm\beta \in \mathbb{R}^m are estimated using sparse linear regression (the Lasso):

    β^v=arg⁡min⁡β∈Rm12n∥xv−Yβ∥22+α∥β∥1\hat{\beta}_v = \arg\min_{\beta \in \mathbb{R}^m} \frac{1}{2n} \|x_v - Y \beta\|_2^2 + \alpha \|\beta\|_1

    where α>0\alpha > 0 controls the sparsity of the receptive field, and ∥β∥1=∑j=1m∣βj∣\|\beta\|_1 = \sum_{j=1}^m |\beta_j|.

    The regularization strength α\alpha is determined automatically via cross-validation using the Least Angle Regression (LARS) algorithm (LassoLarsCV). Reshaping the estimated coefficient vector β^v\hat{\beta}_v back into the 2D stimulus grid geometry (10×1010 \times 10) yields the spatial receptive field of voxel vv. Neighboring voxels in visual cortex yield receptive fields in neighboring pixel coordinates, recovering the retinotopic layout.

  4. Knowl 4 — Performance of Sparse and Dense Classifiers in Binary Visual Image Decoding

    data/table

    The table below presents the 5-fold cross-validation classification accuracy (mean ±\pm standard deviation) for reconstructing 10×1010 \times 10 binary stimulus pixels from visual cortex fMRI BOLD activity across varying regularization strengths CC:

    Classifier C=0.0005C=0.0005 C=0.001C=0.001 C=0.005C=0.005 C=0.01C=0.01 C=0.05C=0.05 C=0.1C=0.1
    ℓ1\ell_1 Logistic Regression 0.50±0.020.50 \pm 0.02 0.50±0.020.50 \pm 0.02 0.57±0.130.57 \pm 0.13 0.63±0.110.63 \pm 0.11 0.70±0.12\mathbf{0.70 \pm 0.12} 0.70±0.120.70 \pm 0.12
    ℓ2\ell_2 Logistic Regression 0.60±0.110.60 \pm 0.11 0.61±0.120.61 \pm 0.12 0.63±0.130.63 \pm 0.13 0.63±0.130.63 \pm 0.13 0.64±0.130.64 \pm 0.13 0.64±0.13\mathbf{0.64 \pm 0.13}
    ℓ1\ell_1 SVM (SVC) 0.50±0.060.50 \pm 0.06 0.55±0.120.55 \pm 0.12 0.69±0.110.69 \pm 0.11 0.71±0.12\mathbf{0.71 \pm 0.12} 0.69±0.120.69 \pm 0.12 0.68±0.120.68 \pm 0.12
    ℓ2\ell_2 SVM (SVC) 0.67±0.120.67 \pm 0.12 0.67±0.12\mathbf{0.67 \pm 0.12} 0.67±0.120.67 \pm 0.12 0.66±0.120.66 \pm 0.12 0.65±0.120.65 \pm 0.12 0.65±0.120.65 \pm 0.12

    The comparison shows that tuning the inverse regularization parameter CC is critical: overly small values of CC lead to severe underfitting, degrading performance to chance (0.500.50). Sparse ℓ1\ell_1-penalized models (both ℓ1\ell_1 Logistic Regression and ℓ1\ell_1 SVM) attain higher peak reconstruction accuracy (0.70−0.710.70 - 0.71) than dense ℓ2\ell_2-regularized models (0.64−0.670.64 - 0.67), demonstrating the efficacy of sparsity when only a small subset of cortical voxels carries predictive signal for any given pixel.

  5. Knowl 5 — Spatially Constrained Functional Brain Parcellation via Ward Hierarchical Agglomeration

    algorithm

    To group fMRI voxels into functionally homogeneous and spatially contiguous brain parcels, bottom-up agglomerative hierarchical clustering with Ward's criterion is constrained by an image neighborhood graph:

    Input: Preprocessed fMRI data matrix X∈RT×VX \in \mathbb{R}^{T \times V} (TT time points, VV in-mask voxels), 3D binary brain mask MM, target parcel count KK.
    Output: Parcel assignment vector c∈{1,…,K}Vc \in \{1, \dots, K\}^V.
    1. Construct an adjacency graph G=(V,E)G = (V, E) based on the 3D spatial grid neighborhood structure (e.g., 6- or 26-connectivity) restricted to voxels where M=1M = 1.
    2. (Optional) Apply Principal Component Analysis (PCA) along the temporal dimension of XX to reduce dimensionality from TT to dd (d≪Td \ll T) while preserving second-order voxel covariance statistics.
    3. Initialize each voxel v∈Vv \in V as a singleton cluster Cv={v}C_v = \{v\}.
    4. while current number of clusters >K> K do
        Identify the pair of adjacent clusters (A,B)(A, B) connected by an edge in GG that minimizes the increase in total within-cluster variance:
        ΔVar(A,B)=∣A∣⋅∣B∣∣A∣+∣B∣∥xˉA−xˉB∥22\Delta \text{Var}(A, B) = \frac{|A| \cdot |B|}{|A| + |B|} \|\bar{x}_A - \bar{x}_B\|_2^2
        where xˉA,xˉB\bar{x}_A, \bar{x}_B are the mean temporal profiles of clusters AA and BB.
        Merge cluster AA and cluster BB into a new cluster CAB=A∪BC_{AB} = A \cup B.
        Update graph GG by connecting CABC_{AB} to all spatial neighbors of AA and BB.
    5. end while
    6. Return the cluster label cvc_v for each voxel v∈Vv \in V.

    Restricting cluster merges strictly to adjacent nodes in the spatial connectivity graph guarantees that all resulting functional parcels are spatially contiguous.

  6. Knowl 6 — Resting-State Network Extraction via Temporal Concatenation Spatial ICA

    model/method

    Independent Component Analysis (ICA) decomposes resting-state fMRI multivariate signals into spatially independent resting-state functional networks (such as the default mode network) and artifact components by maximizing non-Gaussianity.

    In multi-subject resting-state fMRI analysis using concatenation ICA (Concat-ICA):

    1. Data Preprocessing: Each subject's masked 2D data matrix Xs∈RTs×VX_s \in \mathbb{R}^{T_s \times V} (with TsT_s time points and VV in-mask voxels) is centered and detrended across time to eliminate scanner drifts and linear trends.
    2. Temporal Concatenation: Individual subject matrices are concatenated vertically along the temporal dimension: Xconcat=[X1X2⋮XS]∈R(∑s=1STs)×VX_{\text{concat}} = \begin{bmatrix} X_1 \\ X_2 \\ \vdots \\ X_S \end{bmatrix} \in \mathbb{R}^{\left(\sum_{s=1}^S T_s\right) \times V}
    3. Spatial ICA Decomposition: FastICA is fitted to the transposed matrix Xconcat⊤∈RV×(∑s=1STs)X_{\text{concat}}^\top \in \mathbb{R}^{V \times \left(\sum_{s=1}^S T_s\right)}, treating voxels as samples and concatenated time points as features. Decomposing into KK components yields spatial activation maps S∈RK×VS \in \mathbb{R}^{K \times V} representing shared functional networks across the group.

  7. Knowl 7 — Formatting 4D Neuroimaging Volumes into 2D Data Matrices for Statistical Learning

    model/method

    Standard machine learning algorithms operate on 2-dimensional feature matrices of shape (samples×features)(\text{samples} \times \text{features}), whereas functional neuroimaging data are stored as 4-dimensional spatio-temporal volumes (nx×ny×nz×nt)(n_x \times n_y \times n_z \times n_t) with an associated affine transformation matrix.

    The conversion workflow proceeds as follows:

    1. Spatial Resampling: The 3D/4D scan is resampled onto a common template affine grid (e.g., MNI coordinates) or downsampled (e.g., 2 mm or 3 mm isotropic resolution) via affine interpolation.
    2. Signal Cleaning: Time series undergo detrending to eliminate linear baseline drift, variance normalization (scaling voxel time series to unit variance), and temporal filtering (via Fourier transforms or Butterworth bandpass filters to isolate physiological BOLD frequencies).
    3. Boolean Masking: A 3D binary brain mask M∈{0,1}nx×ny×nzM \in \{0, 1\}^{n_x \times n_y \times n_z} containing VV active voxels is applied to discard out-of-brain voxels carrying noise. Extracting in-mask voxels yields a 2D matrix X∈Rnt×VX \in \mathbb{R}^{n_t \times V} (where samples are time points/trials and features are voxels) for supervised decoding/encoding, or X∈RV×ntX \in \mathbb{R}^{V \times n_t} for spatial decomposition (spatial ICA and spatial parcellation).
    4. Unmasking: Model outputs (such as classifier feature weights or cluster labels of size VV) are mapped back into the original 3D volume shape using MM for visualization.

  8. Knowl 8 — Spatial Contiguity Differences Between Spatially Constrained Ward Parcellation and K-Means

    empirical result

    When parcellating resting-state fMRI data into functional regions:

    • KK-Means Clustering: Minimizes overall variance by iteratively assigning voxels to the nearest cluster centroid based on functional time series similarity. Because KK-means does not incorporate spatial neighborhood topology, it produces clusters of voxels scattered across disjoint anatomical regions. Consequently, requesting KK clusters yields a much larger number of fragmented, non-contiguous sub-regions unless intense spatial smoothing is applied.
    • Spatially Constrained Ward Agglomeration: Enforces cluster merges solely between spatially neighboring voxels via an image adjacency graph. This guarantees that every cluster forms a strictly contiguous parcel. Spatially constrained Ward agglomeration is especially effective at higher cluster counts (e.g., K=1,000K = 1{,}000), where it accurately delineates fine functional anatomical boundaries (such as the calcarine sulcus).

Coverage note — None was omitted; the full methodological and empirical workflow of the paper—encompassing data preparation, decoding (SVM/feature selection), visual stimulus reconstruction, encoding/pRF modeling (Lasso/Ridge), resting-state Concat-ICA, and functional parcellations (Ward/K-Means)—has been covered.

References

  1. 1.Beckmann, C. F., and Smith, S. M. (2004). Probabilistic independent component analysis for functional magnetic resonance imaging. Trans. Med. Imaging 23, 137–152. doi: 10.1109/TMI.2003.822821
  2. 2.Biswal, B., Zerrin Yetkin, F., Haughton, V., and Hyde, J. (1995). Functional connectivity in the motor cortex of resting human brain using echo-planar MRI. Magn. Reson. Med. 34, 53719. doi: 10.1002/mrm.1910340409
  3. 3.Calhoun, V. D., Adali, T., Pearlson, G. D., and Pekar, J. J. (2001). A method for making group inferences from fMRI data using independent component analysis. Hum. Brain Mapp. 14, 140. doi: 10.1002/hbm.1048
  4. 4.Craddock, R., James, G., Holtzheimer III, P., Hu, X., and Mayberg, H. (2011). A whole brain fmri atlas generated via spatially constrained spectral clustering. Hum. Brain Mapp. 33, 1914–1928. doi: 10.1002/hbm.21333
  5. 5.Detre, G., Polyn, S., Moore, C., Natu, V., Singer, B., Cohen, J., et al. (2006). “The multi-voxel pattern analysis (mvpa) toolbox,” in Poster Presented at the Annual Meeting of the Organization for Human Brain Mapping (Florence, Italy). Available online at: http://www.csbmb.princeton.edu/mvpa
  6. 6.Efron, B., Hastie, T., Johnstone, L., and Tibshirani, R. (2004). Least angle regression. Ann. Stat. 32, 407–499. doi: 10.1214/009053604000000067
  7. 7.Friston, K. (2007). Statistical Parametric Mapping: The Analysis of Functional Brain Images. London: Academic Press.
  8. 8.Gorgolewski, K., Burns, C. D., Madison, C., Clark, D., Halchenko, Y. O., Waskom, M. L., et al. (2011). Nipype: a flexible, lightweight and extensible neuroimaging data processing framework in python. Front. Neuroinform. 5:13. doi: 10.3389/fninf.2011.00013
  9. 9.Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., and Witten, I. H. (2009). The weka data mining software: an update. ACM SIGKDD Explor. Newsl. 11, 10–18. doi: 10.1145/1656274.1656278
  10. 10.Hanke, M., Halchenko, Y. O., Sederberg, P. B., Hanson, S. J., Haxby, J. V., and Pollmann, S. (2009). PyMVPA: a python toolbox for multivariate pattern analysis of fMRI data. Neuroinformatics 7, 37–53. doi: 10.1007/s12021-008-9041-y
  11. 11.Hanson, S. J., and Halchenko, Y. O. (2008). Brain reading using full brain support vector machines for object recognition: there is no “face” identification area. Neural Comput. 20, 486–503. doi: 10.1162/neco.2007.09-06-340
  12. 12.Hanson, S. J., Matsuka, T., and Haxby, J. V. (2004). Combinatorial codes in ventral temporal lobe for object recognition: Haxby (2001) revisited: is there a “face” area? Neuroimage 23, 156–166. doi: 10.1016/j.neuroimage.2004.05.020
  13. 13.Hastie, T., Tibshirani, R., and Friedman, J. J. H. (2001). The Elements of Statistical Learning, Vol. 1. New York, NY: Springer. doi: 10.1007/978-0-387-21606-5
  14. 14.Haxby, J. V., Gobbini, I. M., Furey, M. L., Ishai, A., Schouten, J. L., and Pietrini, P. (2001). Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science 293, 2425. doi: 10.1126/science.1063736
  15. 15.Hunter, J. D. (2007). Matplotlib: a 2d graphics environment. Comput. Sci. Eng. 9, 90–95. doi: 10.1109/MCSE.2007.55
  16. 16.Hyvärinen, A., and Oja, E. (2000). Independent component analysis: algorithms and applications. Neural networks 13, 411–430. doi: 10.1016/S0893-6080(00)00026-5
  17. 17.Kiviniemi, V., Kantola, J., Jauhiainen, J., Hyvärinen, A., and Tervonen, O. (2003). Independent component analysis of nondeterministic fmri signal sources. Neuroimage 19, 253. doi: 10.1016/S1053-8119(03)00097-1
  18. 18.Kriegeskorte, N., Goebel, R., and Bandettini, P. (2006). Information-based functional brain mapping. Proc. Natl. Acad. Sci. U.S.A. 103, 3863–3868. doi: 10.1073/pnas.0600244103
  19. 19.Michel, V., Gramfort, A., Varoquaux, G., Eger, E., Keribin, C., and Thirion, B. (2012). A supervised clustering approach for fMRI-based inference of brain states. Pattern Recogn. 45, 2041. doi: 10.1016/j.patcog.2011.04.006
  20. 20.Millman, K. J., and Brett, M. (2007). Analysis of functional magnetic resonance imaging in python. Comput. Sci. Eng. 9, 52–55. doi: 10.1109/MCSE.2007.46
  21. 21.Miyawaki, Y., Uchida, H., Yamashita, O., Sato, M.-A., Morito, Y., Tanabe, H. C., et al. (2008). Visual image reconstruction from human brain activity using a combination of multiscale local image decoders. Neuron 60, 915–929. doi: 10.1016/j.neuron.2008.11.004
  22. 22.Mur, M., Bandettini, P. A., and Kriegeskorte, N. (2009). Revealing representational content with pattern-information fMRI–an introductory guide. Soc. Cogn. Affect. Neurosci. 4, 101–109. doi: 10.1093/scan/nsn044
  23. 23.Naselaris, T., Kay, K. N., Nishimoto, S., and Gallant, J. L. (2011). Encoding and decoding in fMRI. Neuroimage 56, 400–410. doi: 10.1016/j.neuroimage.2010.07.073
  24. 24.O’Toole, A. J., Jiang, F., Abdi, H., Pénard, N., Dunlop, J. P., and Parent, M. A. (2007). Theoretical, statistical, and practical perspectives on pattern-based classification approaches to the analysis of functional neuroimaging data. J. Cogn. Neurosci. 19, 1735–1752. doi: 10.1162/jocn.2007.19.11.1735
  25. 25.Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., et al. (2011). Scikit-learn: machine learning in python. J. Mach. Learn. Res. 12, 2825.
  26. 26.Pereira, F., Mitchell, T., and Botvinick, M. (2009). Machine learning classifiers and fMRI: a tutorial overview. Neuroimage 45, S199–S209. doi: 10.1016/j.neuroimage.2008.11.007
  27. 27.Schaul, T., Bayer, J., Wierstra, D., Sun, Y., Felder, M., Sehnke, F., et al. (2010). PyBrain. J. Mach. Learn. Res. 11, 743–746.
  28. 28.Schrouff, J., Rosa, M. J., Rondina, J., Marquand, A., Chu, C., Ashburner, J., et al. (2013). PRoNTo: pattern recognition for neuroimaging toolbox. Neuroinformatics 11, 319–337. doi: 10.1007/s12021-013-9178-1
  29. 29.Smith, S., Fox, P., Miller, K., Glahn, D., Fox, P., Mackay, C., et al. (2009). Correspondence of the brain’s functional architecture during activation and rest. Proc. Natl. Acad. Sci. U.S.A. 106, 13040. doi: 10.1073/pnas.0905267106
  30. 30.Smith, S. M., Jenkinson, M., Woolrich, M. W., Beckmann, C. F., Behrens, T. E., Johansen-Berg, H., et al. (2004). Advances in functional and structural MR image analysis and implementation as FSL. Neuroimage 23, S208–S219. doi: 10.1016/j.neuroimage.2004.07.051
  31. 31.Sonnenburg, S., Rätsch, G., Henschel, S., Widmer, C., Behr, J., Zien, A., et al. (2010). The SHOGUN machine learning toolbox. J. Mach. Learn. Res. 11, 1799.
  32. 32.Thirion, B., Flandin, G., Pinel, P., Roche, A., Ciuciu, P., and Poline, J. (2006). Dealing with the shortcomings of spatial normalization: Multi-subject parcellation of fMRI datasets. Hum. Brain Mapp. 27, 678. doi: 10.1002/hbm.20210
  33. 33.Tibshirani, R. (1996). Regression shrinkage and selection via the Lasso. J. R. Stat. Soc. 58, 267–288.
  34. 34.Van Der Walt, S., Colbert, S. C., and Varoquaux, G. (2011). The numpy array: a structure for efficient numerical computation. Comput. Sci. Eng. 13, 22–30. doi: 10.1109/MCSE.2011.37
  35. 35.Varoquaux, G., and Craddock, R. C. (2013). Learning and comparing functional connectomes across subjects. Neuroimage 80, 405. doi: 10.1016/j.neuroimage.2013.04.007
  36. 36.Varoquaux, G., Gramfort, A., Pedregosa, F., Michel, V., and Thirion, B. (2011). Multi-subject dictionary learning to segment an atlas of brain spontaneous activity. Inf. Process Med. Imaging 22, 562–573. doi: 10.1007/978-3-642-22092-0_46
  37. 37.Varoquaux, G., Sadaghiani, S., Pinel, P., Kleinschmidt, A., Poline, J. B., and Thirion, B. (2010). A group model for stable multi-subject ICA on fMRI datasets. Neuroimage 51, 288. doi: 10.1016/j.neuroimage.2010.02.010

Citation

MLA
Abraham, A., et al. “Machine Learning for Neuroimaging with Scikit-Learn”. arXiv, 2014, http://arxiv.org/abs/1412.3919v1.
APA
Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Muller, A., Kossaifi, J., Gramfort, A., Thirion, B., & Varoquaux, G. (2014). Machine Learning for Neuroimaging with Scikit-Learn. arXiv. http://arxiv.org/abs/1412.3919v1
Chicago
Abraham, A., F. Pedregosa, M. Eickenberg, et al. 2014. “Machine Learning for Neuroimaging with Scikit-Learn”. arXiv. http://arxiv.org/abs/1412.3919v1.
Harvard
Abraham, A. et al. (2014) “Machine Learning for Neuroimaging with Scikit-Learn”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1412.3919v1.
Vancouver
1. Abraham A, Pedregosa F, Eickenberg M, Gervais P, Muller A, Kossaifi J, Gramfort A, Thirion B, Varoquaux G (2014) Machine Learning for Neuroimaging with Scikit-Learn. arXiv

BibTeX

@article{abraham2014machine,
  title = {Machine Learning for Neuroimaging with Scikit-Learn},
  author = {Abraham, Alexandre and Pedregosa, Fabian and Eickenberg, Michael and Gervais, Philippe and Muller, Andreas and Kossaifi, Jean and Gramfort, Alexandre and Thirion, Bertrand and Varoquaux, Gäel},
  year = {2014},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1412.3919v1},
  eprint = {1412.3919}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF