Multi-subject dictionary learning to segment an atlas of brain spontaneous activity
read more
Citations
Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI
Machine learning for neuroimaging with scikit-learn.
Resting-state fMRI in the Human Connectome Project
Local-Global Parcellation of the Human Cerebral Cortex From Intrinsic Functional Connectivity MRI
Sparse Modeling for Image and Vision Processing
References
A Fast Iterative Shrinkage-Thresholding Algorithm for Linear Inverse Problems
Atomic Decomposition by Basis Pursuit
De-noising by soft-thresholding
An information-maximization approach to blind separation and blind deconvolution
Independent component analysis: algorithms and applications
Related Papers (5)
The organization of the human cerebral cortex estimated by intrinsic functional connectivity
Frequently Asked Questions (15)
Q2. How many features are extracted from the mask?
After correction for slice timing differences, motion correction and inter-subject spatial normalization using SPM5, the authors extract the time series on a mask of the brain, resulting in roughly 500 samples per subject, and 25000 features – a dataset of 2Go.
Q3. What is the spectral norm of the Laplacian operator?
The Lipschitz constant of the smooth term is given by 1 + γ‖L‖, where ‖L‖ stands for the spectral norm of the Laplacian operator.
Q4. What is the popular method of learning a dictionary?
In neuroimaging, multi-subject dictionary learning using a fixed group model (ζ = 0) in combination with ICA is popular, and called concatenated ICA [6].
Q5. What is the problem in the dictionary learning framework?
In this paper, the authors formulate the problem in the dictionary learning framework and reject observation noise based on the assumption that the relevant patterns are spatially sparse [10, 26], and the authors focus on the choice of the involved parameters.
Q6. What is the importance of ICA in population studies?
This is important in population studies as, unlike previous work, their approach uses the full multi-subject data set to perform simultaneously denoising and latent factor estimation.
Q7. What is the simplest way to minimize E as a function of Vs?
Following [18], the authors use a block coordinate descent, to minimize E as a function of Us. Solving Eq. (3) as a function of Vs corresponds to a ridge regression problem on the variable (Vs − V)T , the solution of which can be computed efficiently (line 9, algorithm 1).
Q8. Why did the authors measure the likelihood of left-out data?
The authors have measured by 3-fold cross validation the likelihood of left-out data as a function of model order for ICA and SPCA, but because of lack of time and computing resource not for MSDL (see Fig. 3).
Q9. How can the model learn patterns that account for subject-to-subject variability?
In addition, at high model-order the model can learn patterns that account for subject-to-subject variability, as the authors will see on simulated data.
Q10. What is the probability of left-out data?
as already noted in [21], setting a high model order may only lead to a saturation of the likelihood of left-out data, and not a decrease.
Q11. What is the simplest way to minimize E as a function of V?
Minimizing E as a function of V corresponds to minimizing ∑Ss=1 1 2 ‖vs − v‖22 + λ µ Ω(v) for all column vectors v of V. Thesolution is a proximal operator [8], as detailed in lemma 1.Lemma 1. argmin v( ∑Ss=1 1 2 ‖vs−v‖22+γ Ω(v))= prox γ/S
Q12. How do the authors compute the assignment matching the estimated maps with the ground truth?
For each of these datasets, the authors compute the best assignment matching the estimated maps with the ground truth using the Kuhn-Munkres algorithm[20] to maximize crosscorrelation.
Q13. What is the simplest way to solve the optimization problem?
Algorithm 1 Solving optimization problem given in Eq. (3)Input: {Ys ∈ Rn×p, s = 1, . . . , S}, the time series for each subject; k, the number of maps; an initial guess for V. Output: V ∈ Rp×k the group-level spatial maps, {Vs ∈ Rp×k} the subject-specific spatial maps, {Us ∈ Rn×k} the associated time series.
Q14. What is the way to recover subject-level maps?
If there is no spatial jitter across subjects, MSDL and SPCA perform similarly for the recovery of population-level maps, but SPCA outperforms MSDL for the recovery of subject-level maps.
Q15. What is the difference between the two approaches?
Their multi-subject model however differs from generalized canonical correlation analysis [16], and its sparse variants [1], as these approaches do not model subject-specific latent factors and thus do not allow for two levels of variance.