Source-linked AI summary
Fine-grain atlases of functional modes for fMRI analysis
Kamalaker Dadi, Gaël Varoquaux, Antonia Machlouzarides-Shalit, Krzysztof J. Gorgolewski, Demian Wassermann, Bertrand Thirion, Arthur Mensch
TL;DR
Population imaging brings computational and statistical challenges as functional-imaging datasets grow, while reduced signals are important starting points for analysis. The paper introduces DiFuMo, functional-mode atlases trained across studies at multiple resolutions. The authors report strong performance across analysis tasks and identify high-dimensional soft atlases as useful representations of brain activity.
Problem
Population imaging creates large datasets that are challenging to analyze, and existing highly resolved soft functional modes have been limited despite their potential to represent functional gradients.
Source: §1 · p. 1Method
DiFuMo provides functional-mode atlases at 64–1024 resolutions, learned from large-scale fMRI data and labeled by anatomical location.
Source: §1 · p. 2, §2.3 · p. 3, §2.5 · p. 3Results
Takeaways & Limitations
High-dimensional soft functional modes capture rich descriptions of neural activity while supporting a broad set of fMRI studies.
Source: §5 · p. 8, §6 · p. 9DiFuMo includes modes outside gray matter, so its atlas is not restricted to gray-matter structures.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Population imaging markedly increased the size of functional-imaging datasets, shedding new light on the neural basis of inter-individual differences. Analyzing these large data entails new scalability challenges, computational and statistical. For this reason, brain images are typically summarized in a few signals, for instance reducing voxel-level measures with brain atlases or functional modes. A good choice of the corresponding brain networks is important, as most data analyses start from these reduced signals. We contribute finely-resolved atlases of functional modes, comprising from 64 to 1024 networks. These dictionaries of functional modes (DiFuMo) are trained on millions of fMRI functional brain volumes of total size 2.4TB, spanned over 27 studies and many research groups. We demonstrate the benefits of extracting reduced signals on our fine-grain atlases for many classic functional data analysis pipelines: stimuli decoding from 12,334 brain responses, standard GLM analysis of fMRI across sessions and individuals, extraction of resting-state functional-connectomes biomarkers for 2,500 individuals, data compression and meta-analysis over more than 15,000 statistical maps. In each of these analysis scenarii, we compare the performance of our functional atlases with that of other popular references, and to a simple voxel-level analysis. Results highlight the importance of using high-dimensional "soft" functional atlases, to represent and analyse brain activity while capturing its functional gradients. Analyses on high-dimensional modes achieve similar statistical performance as at the voxel level, but with much reduced computational cost and higher interpretability. In addition to making them available, we provide meaningful names for these modes, based on their anatomical location. It will facilitate reporting of results.
1. Introduction
Population imaging creates large, high-dimensional datasets, making compact but meaningful image representations valuable. The paper introduces DiFuMo, a multiresolution set of functional modes, and reports that benchmarked signals from these modes support functional-image analysis.
Motivation
- Population imaging produces terabytes of functional brain images, creating storage and high-dimensional analysis challenges despite revealing neural bases of individual differences.Image-derived phenotypes summarize image information in low-dimensional signals while retaining meaningful brain representations.Source: §1 · p. 1
Research gap
- Most functional atlases use locally uniform parcellations that represent functional gradients poorly, while highly resolved soft functional modes have been unavailable amid computational and statistical challenges.DiFuMo addresses this gap with large-scale training on 3 million volumes totaling 2.4 TB.Source: §1 · pp. 1–2
Contributions
- DiFuMo provides functional modes at 64, 128, 256, 512, and 1024 dimensions, optimized across experimental conditions and benchmarked for functional-image analysis.The authors also provide anatomical labels for the modes to facilitate reporting.Source: §1 · p. 2
2. Methods: data-driven fine-grain functional modes
The methods develop DiFuMo as a scalable functional-mode atlas and describe how its signals are extracted and anatomically labeled. They situate the approach among image-derived phenotypes and functional-atlas methods, then specify the training data, factorization, and signal-extraction procedure.
2.1. Context: Image Derived Phenotypes
- Image-derived phenotypes reduce storage and analysis burdens; voxel-level connectome analysis entails billions of connections, so signals are commonly averaged over regions or networks.Source: §2.1 · p. 2
2.2. Methods for data-driven functional atlases
- Functional atlases can define homogeneous clusters or continuous modes; DiFuMo uses sparse dictionary learning, which scales to high model orders and produces localized functional regions.The factorization represents fMRI time series through spatial modes and temporal loadings.Source: §2.2 · pp. 2–3, Figure 1 · p. 2
2.3. DiFuMo extraction: model and data
- DiFuMo dictionaries with 64–1024 modes were learned from concatenated BOLD data across 2,192 sessions; sparsity and positivity constraints promote localized maps, with λ chosen for approximate whole-brain coverage.Source: §2.3 · p. 3
2.4. Extracting signal on functional modes
- Because DiFuMo modes are continuous and may overlap, signal extraction uses linear regression with the dictionary pseudo-inverse rather than simple averaging.For non-overlapping parcellations, the same regression reduces to averaging image values within each cluster.Source: §2.4 · p. 3
2.5. Region names: relation to anatomical structures
- Each functional mode is named for the anatomical structure it most overlaps, using reference atlases and, when overlap is weak, standard anatomy references.Source: §2.5 · pp. 3–4
3. Brain-image analysis on functional modes
The section benchmarks reduced functional-atlas signals across task-fMRI mapping and decoding, connectome-based prediction, and image reconstruction. Each pipeline compares atlas-derived signals with other atlases and, where described, voxel-level analysis.
3.1. Benchmarking several functional atlases
- The benchmark varies atlas spatial components while using the same signal-extraction function, and includes voxel-level analyses as a higher-cost baseline.Reference atlases include ICA, BASC, Craddock, Gordon, Shirer/Altmann, and Schaefer.Source: §3.1 · p. 4
3.2. Mapping brain response: standard task-fMRI analysis
- Task-fMRI GLM models BOLD signals from voxel-level or reduced inputs; reconstructed beta maps are converted to z-maps, FDR-thresholded, and compared with voxel-level maps using Dice similarity.The RSVP analysis estimates beta maps across six conditions and repeats random selections of three sessions ten times to assess Dice variability.Source: §3.2 · p. 5
3.3. Decoding experimental stimuli from brain responses
- Stimulus decoding uses linear ridge regression or SVC models, testing on held-out sessions or subjects and comparing voxel-level z-maps with atlas-reduced data.The evaluation covers six open-access task-fMRI studies, including inter-subject and intra-subject decoding.Source: §3.3 · p. 5
3.4. Predicting phenotype from functional connectomes
- Phenotype prediction projects BOLD data onto an atlas, computes connectomes, and fits penalized logistic or ridge regression models to tangent-space covariance features.Evaluation uses 20 subject-level train/test splits with a 25% test set across seven datasets and diverse phenotypic targets.Source: §3.4 · pp. 5–6
3.5. Quality of image reconstruction
- Image reconstruction projects statistical maps onto atlases and reconstructs them from loadings; R2 measures retained image variance, while meta-analysis compares averages of voxel-level and reconstructed maps.The reconstruction assessment uses 15,542 NeuroVault maps after excluding maps from DiFuMo training studies and maps failing quality inspection; meta-analysis covers motor, language, and face-recognition terms.Source: §3.5 · p. 6
4. Results: comparing atlases for analyses
The results compare functional atlases across task-fMRI analysis, decoding, connectome-based prediction, and image reconstruction. Higher resolutions tend to help task analyses and reconstruction, while connectome prediction favors intermediate dimensionality.
4.1. Brain mapping: standard task-fMRI analysis
- 1024-dimensional DiFuMo maps correspond more closely to voxel-level GLM maps than 1000-dimensional Schaefer maps, with Dice similarity comparable to voxel-level run-to-run variation.DiFuMo at 512 or 1024 dimensions also gives results close to the voxel-level gold standard; dimension reduction reduces the multiple-comparisons burden.Source: Figure 3 · p. 6, §4.1 · pp. 6–7
4.2. Decoding mental state from brain responses
- Across six task-fMRI experiments, high-resolution DiFuMo (k = 1024) and Schaefer (k = 1000) give the best decoding performance and on average outperform voxel-level prediction.The reconstructed high-dimensional decoding maps are interpretable and capture information similar to voxel-level analysis.Source: Figure 4 · p. 7, §4.2 · p. 7
4.3. Predicting traits from functional connectomes
- DiFuMo k = 256 gives the best connectome-based prediction overall, followed by Craddock k = 400 and BASC k = 444; optimal dimensionality varies by target.A voxel-level baseline is omitted because it would require covariance matrices around 100,000 × 100,000, described as computationally and statistically intractable.Source: Figure 6 · p. 7, §4.3 · p. 7
4.4. Fraction of the original signal captured
- DiFuMo k = 1024 captures 70% of the original voxel-level signal, and its reconstructed meta-analysis maps better capture gradients than discrete parcellations such as BASC.The reconstruction comparison covers 15,542 statistical images; higher-order atlases generally capture more signal.Source: Figure 7 · p. 8, §4.4 · p. 8
5. Discussion
DiFuMos are high-resolution soft functional modes derived from large-scale fMRI data, intended to capture functional gradients while reducing analysis costs. Across analytic tasks, higher dimensionality generally improves performance, while soft modes tend to outperform hard parcellations at a given dimensionality.
5. Discussion
- DiFuMos combine high resolution and soft modes to capture functional gradients, with 1000× storage reduction and 100× computational speed-up for signal analysis.They are derived from BOLD time series across many studies and provide a few hundred to a thousand subdivisions.Source: §5 · p. 8
- Analytic gains generally increase with atlas dimensionality, except functional-connectome studies, where performance is optimal around 200 nodes; soft modes tend to outperform hard parcellations at matched dimensionality.Choosing node count trades representation complexity against analytic performance.Source: §5 · p. 8
- Training on 2,192 sessions across 27 studies yields modes adapted to EPI signals; large datasets also produce more spatially regular components than training on less data.The modes include white matter and CSF regions, not only gray matter.Source: §5 · p. 8, Figure 8 · p. 9
- Modes outside gray matter can help separate noise or support global-signal removal, while practitioners can restrict extraction to a gray-matter mask when appropriate.Some modes segment ventricles or interfaces.Source: §5 · pp. 8–9
- Sparsity and non-negativity yield localized, anatomically interpretable modes, while increasing dimensionality progressively subdivides distributed networks into smaller regions.For example, the putamen subdivisions separate along the anterio-posterior direction and then by hemisphere.Source: §5 · p. 9, Figure 8 · p. 9
6. Conclusion
The paper provides multidimensional atlases of soft functional modes for extracting functional signals. These atlases perform well across diverse analytic tasks and are intended to ease computational demands and support interpretation across studies.
6. Conclusion
- The authors provide multidimensional functional-mode atlases for extracting functional signals, with modes labeled by the neuroanatomical structures they contain.The atlases draw on more than 2,000 sessions across 27 studies and 2.4TB of data.Source: §6 · p. 9
- DiFuMos perform well across GLM analysis, mental-process decoding, and functional-connectivity analysis, while reduced signals lower computational burdens.Shared functional modes across studies can facilitate comparison and interpretation.Source: §6 · p. 9
- The atlases capture rich descriptions of neural activity and are presented as suitable for a broad set of fMRI studies.The paper describes them as the only high-dimensional soft functional modes available.Source: §6 · p. 9
1/20th of full training size
The appendix compares components trained on different data sizes and reports computational benefits from using DiFuMo-reduced signals. It also points to supplementary details on the stimulus-decoding pipeline and its datasets.
1/20th of full training size
- Components trained on the full input set have more spatial regularity than those trained on 100 volumes, which show more overlap in some brain regions.The comparison concerns 1024 components.Source: Figure A2 · p. 13, Appendix A · p. 13
Appendix B. Details on stimulus decoding
- The decoding appendix supplies further pipeline details and summarizes datasets and prediction tasks for six decoding tasks.It also notes that HCP and IBC z-maps were computed using GLMs, while NeuroVault provided β-maps for Emotion and Pain.Source: Appendix B · p. 13, Table A2 · p. 14
Appendix A. Performance of DiFuMos
- On HCP high-resolution images, predicting mental state with DiFuMo 1024 features reduces computational time by a factor of 200 versus voxel-level features.Table A1 averages time over 20 cross-validation folds across several task-fMRI conditions.Source: Table A1 · p. 14
Appendix B.1. Input data and pre-processing pipelines
The decoding pipeline uses unthresholded statistical maps as inputs, drawn from NeuroVault, HCP, and IBC. It compares atlas-based feature representations, including DiFuMos, against voxel-level decoding.
Appendix B.1. Input data and pre-processing pipelines
- The pipeline classifies unthresholded statistical maps; HCP and IBC maps are GLM-derived z-maps, while NeuroVault supplies β-maps for Emotion and Pain.The dataset and prediction-task overview covers six decoding tasks.Source: Appendix B.1 · p. 14
- For NeuroVault Emotion and Pain maps, ridge regression predicts emotion ratings and Linear SVC predicts pain level across three classes.The maps are downloaded from collections 503 and 504.Source: Appendix B.1 · p. 14
- Decoding from high-order dictionaries, especially DiFuMos, performs similarly or better than decoding from voxels.Figure A3 reports mean prediction scores with standard-deviation error bars for each atlas and target.Source: Figure A3 · p. 14
- HCP z-maps from working-memory, gambling, and relational-processing tasks are classified with Linear SVC to predict psychological conditions.The task labels define either two-class or four-class prediction.Source: Appendix B.1 · pp. 14–15
- IBC motor decoding classifies left versus right button presses within each subject, training on all but one session and testing on the held-out session.Each of 13 subjects contributes 80 z-maps across four sessions.Source: Appendix B.1 · p. 15
- The GLM uses stimulus-timing regressors convolved with hemodynamic models and polynomial terms for low-frequency drifts.The specification applies to z-maps estimated from raw fMRI data for both HCP and IBC.Source: Appendix B.1 · p. 15
Appendix B.2. Detailed results
The appendix reports that higher-dimensional functional atlases improve decoding performance and describes a biomarker-prediction evaluation spanning diverse targets. DiFuMo classification maps are also presented as more interpretable than voxel-level maps.
Appendix B.2. Detailed results
- With 1024-atlas decoding, prediction accuracy matches or exceeds voxel-based prediction.The classification-weight maps are smoother, with clearly visible blobs.Source: Appendix B.2 · p. 15
Appendix C. Details on biomarker prediction
- Biomarker prediction considers multiple datasets to cover diverse prediction targets.The datasets, prediction groups, and targets are summarized in Table A3.Source: Appendix C · p. 15
Appendix C.1. Input data and prediction settings
This appendix describes the datasets and temporal preprocessing used for biomarker prediction and reports prediction comparisons across atlases. DiFuMo-based predictions are on par with several reference atlases, and atlas-based maps also illustrate anatomical interpretability.
Appendix C.1. Input data and prediction settings
- Functional-atlas connectome features predict neurodegenerative and neuropsychiatric outcomes, drug-abuse impact, and psychological traits.Source: Appendix C.1 · p. 15
- DiFuMo weight maps for HCP working-memory face decoding clearly delineate the fusiform gyrus and lateral occipital cortex.Source: Figure A4 · p. 15
- The resting-state fMRI prediction pipeline uses logistic regression for dataset groups and ridge regression for CamCAN age prediction.Table A3 identifies the datasets and targets used in the pipeline.Source: Table A3 · p. 16
Appendix C.3. Detailed results
- Up to k = 256 ROIs, BASC and DiFuMo-based atlases give good prediction scores; DiFuMo predictions are on par with UKBB ICA, Craddock et al. (2012), and BASC.Figure A5 markers show mean performance and error bars show standard deviations.Source: Figure A5 · p. 16, Appendix C.3 · p. 17
Appendix C.1. Input data and prediction settings
- Participants with severe scanning artifacts, head movement above 2 mm, or multiple clinical diagnoses were excluded from the analysis.Source: Appendix C.1 · p. 17
- Preprocessing regresses out 10 CompCor components and six motion signals for ADNI, ADNIDOD, COBRE, and CamCAN, while ABIDE, ACPI, and HCP receive no additional preprocessing.Across datasets, regional signals are normalized, detrended, and bandpass-filtered between 0.01 and 0.1 Hz.Source: Appendix C.1 · p. 17
Appendix D. Intra-subject encoding
The intra-subject encoding analysis tests whether atlas-derived GLM estimates predict responses in held-out sessions. Higher-order DiFuMo atlases retain explained variance close to voxel-level analysis while reducing computational cost.
Appendix D. Intra-subject encoding
- A leave-one-session-out evaluation compares observed single-session time series with reconstructions using average β-maps from five training sessions.The analysis fits separate GLMs across acquisition sessions for each subject.Source: Appendix D · p. 17
- Higher-order atlases better explain an unseen session in the language-task protocol, with similar results reported across subjects.Figure A6 illustrates the comparison for one subject.Source: Figure A6 · p. 17
- R2 maps report the proportion of voxel activation variance explained, and the validation compares atlas-derived and voxel-derived maps using correlation and slope.The slope reflects signal loss from atlas projection, which may also reduce noise.Source: Appendix D · pp. 17–18
- DiFuMo with k = 1024 loses only 6% of explained variance R2 versus voxel-level analysis, while lower-order atlases incur greater signal loss.The R2 maps are described as qualitatively comparable, and the 1024-mode atlas is considered suitable for less costly intra-subject encoding.Source: Appendix D · p. 18
Appendix E. Extra meta-analysis maps
The additional meta-analysis maps compare DiFuMo and BASC reductions with voxel-level images, while the naming appendix describes anatomical labels for DiFuMo modes. The appendices also document finer putamen segmentation at higher atlas dimensions and the training corpus.
Appendix E. Extra meta-analysis maps
- For language and face meta-analysis, DiFuMo maps are easier to interpret than BASC maps and show better image-compression performance.The associated figure compares DiFuMo at k = 1024, BASC at k = 444, and voxel-level images.Source: Appendix E · p. 18
Appendix F. DiFuMos naming details
- DiFuMo components are named by anatomical overlap, with spatial qualifiers and hemisphere labels where multiple components map to a region.Components overlapping several anatomical regions are named by a trained neuroanatomist.Source: Appendix F · p. 18
- At higher dimensions, DiFuMo separates left and right putamen modes and represents organization within substructures.The figure notes that this finer segmentation may be crucial in discriminative tasks.Source: Figure A8 · p. 19
- The training atlas corpus totals 2.4 TB, and the appendix tables report its OpenNeuro studies and acquisition parameters.Source: Table A4 · p. 20, Table A5 · p. 21