Brain-imaging studies often summarize a whole-brain recording as a small set of recurring spatial patterns, or "modes": regions that tend to rise and fall together over time. Sparse dictionary learning is a popular way to find them. It factors the data into a few spatial maps, each mostly zero, plus a time course for each map. The method assumes that a given voxel coordinate means the same thing in every person. It does not. After standard alignment to a common anatomical template, functional regions still sit millimeters to centimeters apart across people. The factorization then has two options. It can blur the displaced copies of a region into one broad map, or it can spend extra components describing the same region at slightly different positions. With hundreds of subjects the damage is partly averaged away. With the 20 to 50 subjects typical of clinical sites, the misalignment competes directly with the small true effects.
We propose to learn the group dictionary and a smooth, low-dimensional deformation for each subject together, in one objective. Each subject's differences from the group are represented as a single geometric displacement shared by all modes, not as free per-voxel edits to each map. The bet is about sample efficiency. A shared warp has a few thousand parameters per subject, while free per-map edits have hundreds of thousands. The shared warp should therefore let a small cohort estimate stable, sharper modes. It should also make a pretrained dictionary transferable to a new site by fitting only the new subjects' warps.
Core Claim
The central hypothesis is that part of the dictionary instability and loss of cross-site generalization seen in small cohorts comes from inter-subject functional misalignment. Modeling that misalignment as a shared geometric warp inside the dictionary objective should recover it more efficiently than the per-subject map flexibility existing methods already have.
The claim makes three testable predictions:
- At fixed small N, atoms (the individual spatial maps) should be more reproducible across split halves.
- The number of components at which cross-site prediction peaks should move upward. Components that were "shifted duplicates" of one region stop being needed.
- Warps learned only from resting-state data should improve cross-subject agreement on held-out task data. That can only happen if the warps capture real functional displacement rather than noise.
The closest prior work treats warping and dictionary learning as separate stages. ICA-fNORM takes resting-state group-ICA networks as templates for nonlinear normalization, then transfers the warps to an auditory oddball task. The number of significant voxels in the left middle temporal gyrus rose by over 116%. This result motivates us: it shows that functionally derived warps carry real signal. The warps, however, are computed once, against fixed ICA networks that cannot overlap. Bayesian functional registration aligns univariate activation maps with MCMC. It succeeded on all 33 subjects, where diffeomorphic demons registration failed on 7. It does not scale to whole-brain multivariate decompositions. Connectivity-based cortical alignment warps dense connectivity profiles compressed by singular value decomposition (SVD) and raised inter-subject correlation from 0.072 to 0.085, on 10 subjects.
On the dictionary side, three methods learn on frozen anatomical coordinates: TV-MSDL, DiFuMo and MSDL in the ABIDE biomarker study. DICCCOL-guided sampling avoids warping by sampling signals at landmarks derived from diffusion MRI fiber tracking, the step its authors identify as the bottleneck. It also assumes structural landmarks mark functional hubs.
None of these methods lets the dictionary and the deformation inform each other. Theoretical work on sample complexity and identifiability of dictionaries assumes a perfectly registered grid. We relax that assumption empirically, and in simulation with known ground truth.
Research Program
Let $X_i$ be subject $i$'s time-by-voxel data matrix and $V$ the $k$-by-voxel matrix of group spatial maps. Let $\phi_i$ be a smooth displacement field for subject $i$, parameterized on a coarse B-spline grid. $V\circ\phi_i$ denotes the group maps resampled into subject $i$'s space, and $U_i$ holds that subject's time courses. We solve
$$\min_{V,\{U_i,\phi_i\}} \sum_i \|X_i - U_i\,(V\circ\phi_i)\|_F^2 + \lambda\|V\|_1 + \mu\,\mathrm{TV}(V) + \gamma\sum_i \mathcal{R}(\phi_i) \quad \text{s.t.}\ V\ge 0,\ \textstyle\sum_i \phi_i = 0.$$
The terms mean the following:
- $\|V\|_1$ makes the maps sparse.
- $\mathrm{TV}(V)$ is total variation, which encourages each map to be made of piecewise-constant, spatially compact regions.
- $\mathcal{R}$ is a bending-energy penalty that keeps warps smooth.
- The zero-mean constraint on displacements stops the template drifting toward any single subject. This addresses the reference-selection bias that affects existing functional registration.
The sparsity, total-variation and positivity terms are reused directly from TV-MSDL. We keep its random-subject-subset updates, its proximal solver for $V$ and its duality-gap stopping rule. The only new block is a gradient step on each $\phi_i$, which is cheap because warping $k$ maps is a sparse resampling.
The key design choice concerns subject variability. TV-MSDL and MSDL already let each subject have its own maps, penalized toward the group maps. Here the flexibility is instead one geometric deformation shared by all $k$ modes. With about 1,000 control points this is roughly $3\times10^3$ parameters per subject. Free subject maps would need about $k$ times 50,000. Because far fewer quantities are estimated from the same $T$ time points, the subject-level fit should be much less noisy at small N. We will make this optional: a hybrid variant adds a small free residual per map on top of the warp. That lets the experiments show whether geometry explains most of the subject variability or only some of it.
We will build four things:
- A NumPy/PyTorch solver on MNI volumes.
- A SimTB-style simulator, following the synthetic benchmark used by assisted dictionary learning. It overlapping modes and applies known subject-specific displacement fields at controlled magnitudes.
- A duplicate-atom diagnostic. It counts pairs of atoms whose correlation rises above 0.7 after a small rigid translation of one of them. This operationalizes the "shifted copies" artifact.
- A transfer mode. A dictionary learned at source sites stays frozen, and only the $\phi_i$ and $U_i$ of new-site subjects are fitted.
Baselines
The naive reference is a fixed anatomical atlas, Harvard-Oxford. We extract region time series and apply the tangent-space connectivity embedding with an L2-regularized classifier used in the ABIDE biomarker study. No learning of maps and no functional alignment are involved. It shows what dictionary learning buys at all.
The strongest existing method is MSDL followed by that same tangent-space pipeline, as in the ABIDE biomarker study. That paper reports 67% inter-site ASD classification accuracy on 871 subjects from 17 sites, against about 60% for earlier approaches. It is the strongest competitor for two reasons. It already gives each subject its own maps, and the study found that the map-extraction step was the largest single driver of accuracy. We also run TV-MSDL under the same protocol. No classification accuracy for TV-MSDL was found on Lacuna.
For fairness, every method uses:
- the same preprocessed ABIDE volumes;
- the same leave-site-out folds;
- the same $k$ grid, from 20 to 400;
- the same downstream classifier;
- an equal hyperparameter search budget, tuned by nested inner-site cross-validation;
- equal wall-clock budgets on identical hardware.
Beating MSDL means a mean leave-site-out accuracy gain whose confidence interval over site folds excludes zero, and that holds in the small-site subsampling regime described below. A gain in the full 871-subject regime alone would not suffice.
We add three reference points:
- Two-stage. Warps are first estimated against fixed MSDL group maps, in the spirit of ICA-fNORM. Then MSDL is rerun on the warped data.
- Pretrained atlas. Data are projected onto pretrained DiFuMo modes at $k=64$ to 1024. DiFuMo reports that $k=1024$ reconstructs about 70% of voxel-level signal.
- Federated reference. Fed-Align reports 67.6% at NYU, against 60.1% for a single-site model. It used four sites and a Harvard-Oxford atlas, so it is cited only as context, not as a matched comparison.
The key ablation replaces the shared warp with free per-subject map deviations under the same small parameter budget, implemented as low-rank per-map residuals. This isolates whether the geometric form of the flexibility is what matters.
Why not just…?
Why not just smooth more heavily? Gaussian smoothing hides misalignment by lowering resolution everywhere. Connectivity-based alignment and Bayesian registration both report that it wipes out focal structure, whereas functional alignment raised peak t-statistics. We include smoothing at 4, 6, 8 and 10 mm full width at half maximum (FWHM) as a baseline sweep. If some kernel matches our split-half reproducibility and accuracy, we will report that the warp adds nothing beyond low-pass filtering.
Why not just let each subject have its own maps, as MSDL already does? This is the real threat. Per-subject maps can in principle absorb a displacement. The ablation above settles it. If free low-rank deviations with the same parameter count match the warp, the geometric prior is unnecessary and we will say so. We expect the warp to win mainly below about 20 subjects per site, and to win on the rest-to-task transfer test. Only a genuine spatial displacement, not a rest-specific map edit, should carry over to task data.
Why not just warp first and then run standard dictionary learning? The two-stage baseline tests this directly. It could be nearly as good. In that case the contribution shrinks to "use dictionary maps, not ICA networks, as registration targets," which is still useful. The joint method should help most where the first-stage template is itself blurred by misalignment, which is the small-N case. We will report the gap as a function of N.
Why not just use a lower $k$ or a pretrained DiFuMo atlas? The ABIDE study found that inter-site prediction peaks at 40 to 100 regions. That can be read as "just use small $k$." Our position is that this peak is partly a misalignment symptom. The prediction is testable: if joint warping leaves both the peak location and the duplicate-atom count unchanged, the prediction fails.
Proposed Experiments
Simulation. First, the SimTB-style simulations with known displacement fields test whether the solver recovers the true maps and warps. We sweep displacement magnitude from 0 to 10 mm and N from 8 to 64. We measure how N-to-recovery curves shift relative to TV-MSDL. This is the empirical counterpart to the identifiability question that existing theory leaves open.
ABIDE, small sites. Second, on ABIDE we subsample training sites to 10, 20 and 40 subjects each, including the 48-subject Leuven site used by TV-MSDL, with repeated draws. Metrics are:
- split-half atom reproducibility;
- held-out explained variance;
- duplicate-atom count as a function of $k$;
- leave-site-out ASD accuracy.
Given that between-group resting-state effects have a median Hedges' $g$ of 0.19 to 0.39 (Small effect size leads to reproducibility failure), we expect modest accuracy gains. We will therefore treat reproducibility and the peak-$k$ shift as the primary endpoints, and classification as secondary.
HCP, rest to task. Third, on HCP subjects with both rest and motor-task runs, we learn warps from rest only. We then apply them to motor-task contrast maps and measure group t-statistics and cross-subject overlap on held-out subjects. For reference, DICCCOL sampling reports a Pearson correlation of about 0.80 between atoms and task stimulus curves on 20 HCP subjects. We will report the same correlation for our dictionary atoms.
New-site transfer. Fourth, for transfer we freeze a dictionary learned on the other ABIDE sites. At each held-out site we fit only the warps using 5, 10 and 20 subjects. We compare with projecting onto the frozen maps without warps and with refitting the full dictionary.
Decision rule. The proposal succeeds only if three results hold together:
- the warp beats the matched-parameter free-deviation ablation at N of 20 or fewer;
- rest-derived warps improve task alignment;
- the duplicate-atom count falls.
If only the ABIDE accuracy moves, we would conclude that the method acts as a regularizer, not as a correction for functional misalignment.
References
- ICA-fNORM: Spatial Normalization of fMRI Data Using Intrinsic Group-ICA Networks
- Bayesian functional registration of fMRI activation maps
- fMRI-Based Inter-Subject Cortical Alignment Using Functional Connectivity
- Extracting Brain Regions from Rest fMRI with Total-Variation Constrained Dictionary Learning
- Fine-grain atlases of functional modes for fMRI analysis
- Deriving reproducible biomarkers from multi-site resting-state data: An Autism-based example.
- A Dictionary Learning Approach for Signal Sampling in Task-Based fMRI for Reduction of Big Data
- The Sample Complexity of Dictionary Learning
- Robust Identifiability in Sparse Dictionary Learning
- Assisted dictionary learning for FMRI data analysis
- Multi-site fMRI analysis using privacy-preserving federated learning and domain adaptation: ABIDE results
- Small effect size leads to reproducibility failure in resting-state fMRI studies