CaliAli: Technical Review and Current Assessment
Review date:4 July 2026
Scope:Nature Communications publication, current CaliAli documentation, public source code, and the latest formal release available at the time of review.
Executive assessment
CaliAli is a longitudinal one-photon calcium-imaging workflow built around a consequential change in processing order. Instead of extracting neuronal components independently from each session and matching those components afterward, CaliAli first prepares and aligns the imaging sessions, concatenates them, and then performs a shared CNMF-E-based extraction. This allows spatial information learned from a stronger session to support recovery of a weaker trace in another session.[^1]
That central idea is real, but its novelty should be described precisely. Aligning and concatenating sessions before extraction was not introduced by CaliAli; the paper itself cites earlier, less-common concatenation-based approaches. CaliAli's more specific contribution is the combination of blood-vessel and neuronal projections for weighted groupwise non-rigid intersession alignment, together with CNMF-E modifications intended to reduce signal dilution and batch-error propagation in long concatenated recordings.[^1]
The published validation program is stronger than that of many longitudinal calcium-imaging tools. It includes known-truth simulations, dual-color anatomical markers, an uninterrupted-recording pseudo-ground-truth experiment, optogenetic tagging, place-cell analyses, and a seven-session dentate-gyrus dataset spanning 99 days.[^1]These experiments support the architecture under the tested conditions. They do not establish that every longitudinal identity is correct, that blood vessels remain reliable in every preparation, or that CaliAli universally outperforms post-extraction tracking systems.
The software is also not a fully automatic black box. Experimental analyses in the paper used manually selected PNR and correlation thresholds, manually defined field-of-view masks, and manual false-positive removal. Current documentation continues to require visual inspection of motion correction and alignment, provides manual component deletion and merging, and allows users to recover missed candidates from residuals.[^1][^4][^5]
As of this review, the latest formal release isv1.4.8.1, released on 29 April 2026. CaliAli remains actively maintained, but it is a MATLAB application with several required proprietary toolboxes and platform-specific compatibility limits.[^7][^9]
The most defensible overall conclusion is:
>CaliAli is a well-motivated and unusually well-validated longitudinal analysis architecture for one-photon calcium imaging. Its strongest contribution is not simply registration or CNMF-E, but the way it couples weighted intersession alignment with joint source extraction. Its main boundaries are alignment uncertainty, manual parameter and curation decisions, activity-dependent detectability, computational growth with session count, and dependence on a tightly coupled MATLAB workflow.
1. What CaliAli actually changes
Most conventional longitudinal pipelines follow this sequence:
session-level extraction
-> field alignment
-> footprint matching
-> longitudinal identity tableCaliAli instead follows the approximate sequence:
per-session preparation
-> intersession alignment
-> session concatenation
-> shared CNMF-E extraction
-> manual and automated reviewThis ordering reduces dependence on a separate post-hoc identity-matching model. After alignment, the software attempts to represent a neuron with one shared spatial component and session-specific temporal activity across the concatenated recording.[^1]
The phrasealign before extractionis nevertheless an abstraction. CaliAli does not apply one registration operation directly to untouched raw files. The current documented workflow first converts or downsamples each recording, performs within-session motion correction or accepts externally corrected input, detrends and scales per-pixel noise, calculates blood-vessel and neuronal projections, aligns the sessions, and only then concatenates them for CNMF-E extraction.[^1][^2][^4]
CaliAli is therefore best described as anend-to-end longitudinal processing system, not merely a cell-registration utility. This distinction has practical consequences: researchers who already possess independently extracted components from another pipeline cannot obtain the full benefit of CaliAli's joint-extraction design without returning toward the imaging data and reprocessing the sessions.
2. Intersession alignment
2.1 Blood vessels are complementary landmarks, not permanent invariants
CaliAli's most distinctive registration feature is the use of blood-vessel projections in addition to activity-dependent neuronal projections. Blood vessels provide elongated and branching structure that can reduce local alignment ambiguity when neighboring somata look similar or when the active neuronal population changes between sessions. CaliAli enhances vessel-like structures with Hessian-based filtering and combines them with neuronal projections based on local correlation and peak-to-noise ratio.[^1]
The vessels should not be described as perfectly stable landmarks. The paper shows that blood-vessel similarity declines with longer intersession intervals and can be affected by focal-plane stability, signal contrast, exposure, and indicator brightness. The authors derived a blood-vessel similarity threshold of 2.7 in their simulation framework; below this threshold, the method switches toward neuron-only alignment and warns the user. In simulations, performance remained robust when vessel visibility was reduced by as much as 50%, then declined as the method transitioned away from vessel information.[^1]
Accordingly, the accurate description is:
>Blood vessels provide an additional anatomical registration signal whose usefulness is measured and can be down-weighted or bypassed when unreliable.
The current vessel score should not be presented as a probability or calibrated confidence value. The implementation creates 100 randomly perturbed non-rigid alignments, compares their vessel-projection correlations with the observed aligned correlation, converts the difference to standard-deviation units, and returns the minimum standardized score across session pairs. The paper defines the same z-like quantity. A score above the chosen threshold is evidence that the observed vessel agreement is stronger than this surrogate distribution; it is not a posterior probability that the registration is biologically correct.[^14]
2.2 Weighted groupwise alignment
CaliAli does not simply warp every session to the first recording or align sessions sequentially. The publication describes a weighted groupwise strategy in which pairwise displacement fields contribute to a common coordinate system. Transformations are weighted using local and global structural similarity so that different parts of the field can be influenced more strongly by sessions that preserve those structures well.[^1]
The currentsessions_non_rigid.msource calculates all session pairs withnchoosek, estimates forward and backward Log-Demons transformations, derives local and global similarity weights, and combines those weighted fields into a transformation for each session.[^10]This all-pairs structure also explains the current documentation's warning that alignment runtime grows approximately quadratically with the number of sessions; the FAQ describes approximately 50 sessions as a usually acceptable range, not a guaranteed limit.[^3]
This is a substantial improvement over choosing one immutable reference session, but it does not remove registration assumptions. Diffeomorphic constraints encourage smooth and continuous transformations; they do not prove that every local warp corresponds to real tissue deformation.
2.3 Within-session and between-session motion are different modules
The draft review treated rigid and non-rigid within-session motion correction as stable current capabilities. Current documentation is inconsistent here:
-the function documentation still describes the non-rigid within-session code as experimental and warns that it can introduce unwanted deformation;
-the current FAQ states that the non-rigid motion-correction module is deprecated and explicitly permits CaImAn or Suite2p as external alternatives.[^3][^12]
This contradiction should be reported rather than silently reconciled. The most conservative current recommendation is to treat CaliAli's rigid within-session correction as the normal integrated path and its non-rigid within-session implementation as deprecated or legacy-experimental. This doesnotapply to the central intersession alignment stage, which still uses non-rigid groupwise registration.
External motion correction is supported, but border artifacts must be removed because black borders can bias intersession alignment. Current documentation asks users to pass externally corrected videos through the downsampling/conversion stage before skipping CaliAli's motion module.[^3]
The integrated dropped-frame repair is also narrower than the phrase may imply. The current helper marks a frame as dropped only when its spatial mean is exactly zero, replaces those frames with missing values, and linearly interpolates along time. It does not detect every corrupted, duplicated, partially written, or abnormally dark frame. A methods section should report how dropped frames were identified and how many were interpolated.[^15]
2.4 Alignment remains a quality-control task
The alignment stage reports a blood-vessel similarity score and spatial-correlation metrics and saves projections for inspection. The documentation explicitly says that users should visually verify alignment before running CNMF-E because apparently successful automated processing can still leave biologically important displacement.[^4]
This is especially important because alignment quality participates directly in the scientific conclusion. Underalignment can make one cell appear to drift or disappear. Overalignment can force distinct structures into apparent correspondence and create artificial stability.
3. Concatenation and customized CNMF-E
3.1 Session boundaries are normalized before concatenation
Sessions can differ in baseline fluorescence, noise amplitude, illumination, and background. If concatenated without correction, the boundary between sessions can become a dominant signal. The paper therefore detrends each pixel and scales its noise distribution before joining sessions. Optional MIN1PIPE-inspired background subtraction and neural enhancement are available for recordings with more complex background fluctuations.[^1]
Current documentation describesneuron_enhance,noise_scale, and a detrending duration among the configurable preprocessing parameters. These should be reported as model choices, not treated as neutral file-conversion settings.[^8]
3.2 Correlation images are calculated from bounded segments
A neuron active during only a small portion of a very long recording can be diluted in a global correlation image. CaliAli addresses this by calculating correlation images from 3,000-frame segments and taking a maximum projection across those segment-level images. In the paper's 50,000-frame demonstration, this preserved the detectability of a neuron active only in the first 20% of the sequence.[^1]
The review should not generalize this into the statement that the software always uses arbitrary short batches or one correlation image per session. The documented publication method specifically uses bounded 3,000-frame segments, while the exact segmentation behavior can depend on the current implementation and input organization.
3.3 Initialization is only partly parallel
Standard CNMF-E initialization is sequential because removing an already initialized source can reveal a weaker nearby source. CaliAli accelerates this by initializingspatially distantcandidates simultaneously, while nearby candidates remain subject to local interaction. It is therefore inaccurate to describe all neuronal initialization as parallel.[^1]
3.4 Unified batch processing
Naive sequential batch CNMF can propagate early mistakes. Later batches inherit previous components and initialize new sources from residuals, which may contain dendritic or other non-somatic structures that satisfy correlation and PNR thresholds. CaliAli instead processes batches internally and combines them into one result. In the publication's formulation:
-background variables are averaged;
-the constant baseline is represented by a minimum projection;
-spatial components are averaged with weights based on squared mean temporal activity in each batch;
-temporal components are concatenated.[^1]
This design allows sessions with stronger activity to contribute more strongly to the estimated footprint. It does not imply that the spatial footprint is literally identical in every session or that large focal changes are harmless.
The current extraction source is also more iterative than a simple one-pass CNMF-E diagram suggests.runCNMFe.mperforms repeated background, spatial, and temporal updates, applies automatic false-positive filtering and multiple merge operations, checks similarity with the preceding iteration, updates residuals, and then enters the final trace-processing stage.[^11]Current manual postprocessing is therefore layered on top of automatic filtering rather than replacing it.
There is a small source-documentation mismatch around dendritic processing. Current parameter documentation enableswith_dendritesby default and disables circular spatial constraints, while theCaliAli_cnmfe.mheader describes selecting a neuron or dendrite-specific route even though the visible dispatcher calls the samerunCNMFeentry point. This does not prove that dendrites are unsupported, but it means publications should report the actual parameters and commit instead of relying on the wrapper's prose description.[^16][^17]
In one simulated 40-session stack, CaliAli achieved an F1 score comparable to full standard CNMF and memory use comparable to sequential batch CNMF. This is useful evidence for the batching design, but it is one defined simulation, not a general proof of fixed-memory processing or universal equivalence to full CNMF.[^1]
3.5 Low-SNR recovery is not recovery from no information
The paper showed that shared extraction improved recovery when a neuron was discernible in one session and low-SNR in another, without reducing high-SNR recovery in the tested simulations.[^1]
This should not be expanded into the claim that CaliAli can recover a completely silent neuron from a session containing no usable signal. The simulation framework required each simulated neuron to exhibit at least one calcium transient per session and used a fixed PNR of 2. Joint spatial information can support a weak trace, but it cannot create temporal evidence absent from the measurement.[^1]
4. Temporal outputs and deconvolution
CaliAli inherits the temporal modeling assumptions of CNMF-E. Spatial components, denoised calcium traces, residual-related signals, and deconvolved activity are model outputs rather than direct measurements of neuronal identity or action potentials.
Current documentation states that the pipeline uses original FOOPSI with an AR(1) model during initialization and matrix-factorization iterations for speed, followed by a final thresholded FOOPSI pass with an AR(2) model. The currentrunCNMFe.msource applies a finalpostprocessDeconvolvedTraces(..., 'foopsi', 'ar2', -5)call after trace detrending and noise scaling.[^5][^11]
These current-software details should be separated from the 2025 paper's broader architectural claims. A methods section should report the release or commit, frame rate, temporal model, thresholding settings, and the exact signal used downstream. DeconvolvedSvalues should be calledinferred event activity, not measured spikes or calibrated firing rates.
5. Human decisions and postprocessing
CaliAli combines automation with consequential user decisions.
The 2025 experimental analyses manually selectedmin_pnrandmin_corrafter inspecting projection images, used manual field-of-view masks, and manually discarded false-positive extractions. The optogenetic experiment deliberately variedmin_corrto demonstrate the precision-recall tradeoff.[^1]
The postprocessing interface does more than sort components by a fixed circularity score. The publication describes a semi-supervised spatial-congruence ranking process: pairwise spatial distances are embedded with multidimensional scaling, manually identified true-positive components define a reference distribution, and Mahalanobis distance is used to rank less-congruent components. Users then delete, merge, or retain components.[^1]
Current documentation also allows users to inspect residuals and manually seed candidates that were missed during the initial pass. Those seeds are sent back through relevant CNMF processing; this is not equivalent to drawing a final ROI and declaring it a neuron.[^5]
These tools are practical strengths, but they create operator dependence. The final component population is not reproducible from automated parameters alone unless manual masks, threshold changes, rejected indices, merges, and residual additions are preserved.
6. What the validation establishes
6.1 Known-truth simulations
The paper simulated non-rigid displacement, different neuron-overlap regimes, changing active populations, signal dilution, and variable SNR. These experiments provide direct evidence because simulated identity and traces are known. Combined vessel and neuronal information outperformed neuron-only registration in the tested alignment scenarios, and CaliAli outperformed CellReg and SCOUT under the reported tracking simulations.[^1]
The comparison still has boundaries. The simulated source shapes and kinetics were sampled from a specific dentate-gyrus dataset, every simulated neuron was required to show at least one transient per session, and the matching definition used spatial and temporal similarities plus one-to-one assignment and a temporal-similarity threshold. The simulation therefore tests the authors' defined domain rather than every possible longitudinal imaging condition.[^1]
The SCOUT comparison also requires a qualification: when SCOUT's native footprint alignment performed poorly in some low-density or remapping simulations, the authors substituted the CellReg alignment module to reduce that confound. The resulting comparison is not simply untouched default SCOUT versus CaliAli.[^1]
6.2 Dual-color near-ground truth
The dual-color analysis used stable nuclear tdTomato to define corresponding cells and evaluated alignment calculated from the activity-dependent green GCaMP channel. It included 218 session pairs from eight mice over 16 days and showed smaller distances between red-channel-defined corresponding ROIs after CaliAli alignment than after the compared methods.[^1]
This is strong evidence foralignment accuracy. It is not a complete ground-truth test of every CNMF-E extraction, temporal trace, manual rejection, or recovered low-SNR event.
6.3 Uninterrupted-recording pseudo ground truth
The authors treated a continuous one-hour recording as a reference, divided it into artificial 5- or 15-minute sessions, reconstructed longitudinal signals, and compared them with the uninterrupted result. CaliAli produced greater similarity to this pseudo ground truth than CellReg or SCOUT.[^1]
This isolates errors introduced by artificial session boundaries, which is valuable. It does not recreate real between-day tissue changes, focal drift, expression change, or biological remapping.
6.4 Optogenetic tagging
Optogenetic response consistency provides a functional identity check: a correctly tracked light-responsive neuron should respond again in another session. CaliAli produced more opto-consistent tracked neurons than the comparison methods in the reported experiments, and artificially increasing misalignment reduced consistency.[^1]
The experiment used sparse expression to improve somatic extraction, and only a minority of extracted components were optogenetically responsive. It is a strong validation of a tagged subset, not a ground-truth label for every detected cell.
6.5 Place-cell analyses
For CA1 place-cell data, concatenated CaliAli processing produced comparable within-session spatial information, neuron count, and position-decoding accuracy to independent processing. Across sessions, the larger tracked population supported better cross-session position decoding under the tested comparisons.[^1]
This is important because it argues against the simple concern that longitudinal consistency was gained by destroying within-session coding. It does not prove that every additional tracked identity was correct.
6.6 Ninety-nine-day dentate-gyrus analysis
CaliAli-tracked populations showed more similar activity vectors across sessions than populations obtained with the comparison methods in a seven-session, 99-day dataset. In the primary population-vector analysis, neurons not detected in every session were excluded to keep vector dimensions equal; setting undetected neurons to zero increased the reported gap.[^1]
The result should therefore be described asgreater method-dependent longitudinal population similarity, not unqualified proof that the dentate-gyrus representation is objectively stable for 99 days. Registration and inclusion rules are part of the measured phenomenon.
7. Software architecture and current status
CaliAli is implemented in MATLAB and currently requires Signal Processing Toolbox, Image Processing Toolbox, Statistics and Machine Learning Toolbox, and Parallel Computing Toolbox. Current documentation lists AVI, M4V, MP4, TIFF, HDF5, and Inscopix ISXD support, while the installation page also lists MKV; codec and conversion support differ by operating system.[^3][^7]
The compatibility documentation is specific rather than universal:
-Windows testing is documented for MATLAB 2022a and 2023a; Windows MATLAB 2024a is listed as unsupported because of an AppDesigner issue.
-The optionalBV_apprequires MATLAB 2023b.
-macOS testing is documented for MATLAB 2024a on Sonoma 14.5, with ISXD limitations on ARM.
-Linux testing is documented for Ubuntu 25.10, with no direct ISXD conversion.[^7]
The current documentation does not define one perfectly consistent default set. The general parameter index listsmin_corr = 0.1,min_pnr = 6, and a within-session non-rigid pyramid of{'BV','neuron','neuron'}. The demo-parameter page listsmin_corr = 0.2,min_pnr = 4, and{'BV','BV','neuron'}. The overview presents six conceptual stages, while the getting-started page numbers five stages and then shows postprocessing in the workflow. The demo page also refers toMIN1PIEbackground subtraction, which appears to be a documentation typo. These discrepancies make the effective options structure, version, and commit more authoritative than a generic statement that defaults were used.[^2][^8][^17][^18]
The latest formal release is v1.4.8.1, a small patch for natural ordering of nested AVI files. The preceding v1.4.8 release added batch downsampling, stronger frame-integrity checks, recovery from partially written intermediates, and CNMF-E startup improvements.[^9]
The repository is licensed under GPL-3.0.[^13]The analysis code is therefore open source, but the required MATLAB runtime and toolboxes are proprietary dependencies.
8. Scaling and resource behavior
CaliAli is memory-aware, not resource-free. Current documentation supports automatic or manually selectedbatch_szvalues for chunked processing. It also warns that CNMF-E is usually the most memory-intensive stage and that automatic memory estimation can fail on some systems, requiring empirical adjustment.[^3][^6]
The alignment algorithm considers all session pairs in the current implementation and consequently scales approximately quadratically with session count. Increasing from 10 sessions to 100 sessions is therefore not merely a tenfold increase in alignment work.[^3][^10]
Disk use also matters. The workflow can create downsampled, motion-corrected, detrended, aligned, checkpoint, and source-extraction outputs. A realistic deployment should benchmark RAM, temporary and final storage, and I/O on representative data rather than relying only on a session-count claim.
9. Principal scientific and operational risks
Registration can influence the biological answer
Longitudinal experiments often ask whether a representation is stable or drifting. Alignment is therefore not neutral preprocessing. Underalignment can produce false drift; over-flexible alignment can produce false stability. Diffeomorphic constraints and multiple validation experiments reduce this risk but do not eliminate it.
Blood-vessel information is preparation-dependent
Vessel projections can be weak, distorted by focal change, or confounded by low contrast. The fallback system is a strength, but users should report the blood-vessel score, whether fallback occurred, whetherForce_BVwas used, and whether an extra neuron-only refinement was applied.[^8]
Shared extraction imposes a spatial-consistency assumption
The weighted shared footprint is advantageous when one biological neuron remains spatially compatible across sessions. Large focal shifts, morphology changes, or imperfect warps can violate that assumption. Activity weighting reduces the influence of weak sessions but does not provide a calibrated identity probability.
Manual choices alter the final population
Thresholds, masks, alignment acceptance, false-positive deletion, merges, and residual recovery all affect cell count and longitudinal identity. These are scientifically legitimate decisions, but they must be recorded.
Comparisons depend on adaptation and evaluation definitions
Some comparator methods were adapted to one-photon projections or supplied with alternative alignment modules. Simulated true positives were defined through the authors' spatiotemporal matching procedure. These choices do not invalidate the paper, but they should accompany claims of superiority.
10. Reproducibility checklist
A defensible CaliAli methods section should record at least:
-CaliAli version, full Git commit, and repository fork;
-MATLAB release, operating system, and toolbox versions;
-input file hashes, session order, acquisition frame rates, dimensions, and codecs;
-spatial and temporal downsampling;
-whether sessions split across files were combined correctly;
-within-session motion-correction software, mode, parameters, the dropped-frame detection rule, interpolated frame indices, and crop;
-detrending window, noise scaling, background subtraction, and non-negative handling;
-blood-vessel size settings, vessel score, projection mode, fallback behavior,Force_BV, andfinal_neurons;
-saved displacement fields and manual alignment acceptance;
-gSig,min_corr,min_pnr, seed mask, patch geometry, batch size, and worker configuration;
-CNMF-E iteration, background, merging, and stopping settings;
-temporal deconvolution method, AR order, event threshold, and signal selected for analysis;
-rejected, merged, and residual-added components;
-reviewer identity and curation criteria;
-checkpoint identities, output hashes, runtime, peak RAM, and storage use.
Current parameter objects, checkpoint files, alignment projections, and release-tagged code provide a useful foundation. They do not automatically capture every manual decision.[^4][^5][^8]
11. Appropriate use
CaliAli is especially compelling when:
-the experiment contains repeated one-photon recordings from the same field;
-active populations change substantially across sessions;
-non-rigid intersession deformation makes footprint-only matching unreliable;
-vessel structure is visible enough to provide complementary landmarks;
-weak-session activity may benefit from a shared spatial model;
-the laboratory is willing to reprocess raw or near-raw recordings in a coordinated MATLAB workflow.
A post-extraction tool such as CellReg may be more practical when recordings have already been processed independently, when raw data are unavailable, or when modular interoperability matters more than joint extraction. CaliAli and post-extraction matching should therefore be viewed as different longitudinal architectures rather than interchangeable implementations.
Overall assessment
CaliAli's most important insight is that long-term neuronal identity cannot always be repaired after independent extraction. By incorporating anatomical context before extraction and sharing spatial information across aligned sessions, the software directly addresses intermittent activity, signal dilution, and non-rigid field changes.
The review should nevertheless avoid turning that architecture into a guarantee. CaliAli does not make blood vessels invariant, recover temporal information that was never recorded, eliminate the need for human curation, or convert hard longitudinal assignments into calibrated biological certainty. Its best evidence is broad and thoughtful, but still preparation- and benchmark-dependent.
The unsupported numerical scorecard in the original draft has therefore been removed. The evidence supports a qualitative conclusion instead:
>CaliAli is a strong longitudinal-first analysis framework with an unusually coherent relationship between registration and source extraction. Its scientific value is highest when the complete workflow, alignment evidence, manual decisions, and uncertainty are treated as part of the reported method rather than hidden preprocessing.
Primary sources
[^1]: Vergara P, Wang Y, Srinivasan S, et al.A comprehensive suite for extracting neuron signals across multiple sessions in one-photon calcium imaging.Nature Communications. 2025;16:3443. https://doi.org/10.1038/s41467-025-58817-z
[^2]: CaliAli current documentation, overview and processing sequence. https://caliali-pv.github.io/CaliAli/latest/
[^3]: CaliAli current FAQ, including memory guidance, approximately quadratic alignment scaling, required visual QC, and deprecated non-rigid within-session motion correction. https://caliali-pv.github.io/CaliAli/latest/FAQ/
[^4]: CaliAli current intersession-alignment documentation. https://caliali-pv.github.io/CaliAli/latest/alignment/
[^5]: CaliAli current signal-extraction and postprocessing documentation. https://caliali-pv.github.io/CaliAli/latest/extraction/
[^6]: CaliAli current low-memory and chunking documentation. https://caliali-pv.github.io/CaliAli/latest/Low_memory/
[^7]: CaliAli installation, dependency, format, and platform-compatibility documentation. https://caliali-pv.github.io/CaliAli/latest/Installation/
[^8]: CaliAli parameter reference. https://caliali-pv.github.io/CaliAli/latest/Parameters_index/
[^9]: CaliAli formal release history. https://github.com/CaliAli-PV/CaliAli/releases
[^10]: Current weighted groupwise registration source,sessions_non_rigid.m. https://github.com/CaliAli-PV/CaliAli/blob/main/Align_sessions/sessions_non_rigid.m
[^11]: Current CNMF-E orchestration source,runCNMFe.m. https://github.com/CaliAli-PV/CaliAli/blob/main/CNMF-e/runCNMFe.m
[^12]: Current non-rigid within-session function documentation, which still labels the module experimental. https://github.com/CaliAli-PV/CaliAli/blob/main/docs/Functions_doc/Non_rigid_mc.md
[^13]: CaliAli GPL-3.0 license and canonical repository. https://github.com/CaliAli-PV/CaliAli
[^14]: Current blood-vessel stability-score implementation,get_BV_NR_score.m. https://github.com/CaliAli-PV/CaliAli/blob/main/Align_sessions/get_BV_NR_score.m
[^15]: Current dropped-frame interpolation helper. https://github.com/CaliAli-PV/CaliAli/blob/main/Motion_Correction/interpolate_dropped_frames.m
[^16]: Current extraction wrapper,CaliAli_cnmfe.m. https://github.com/CaliAli-PV/CaliAli/blob/main/CNMF-e/CaliAli_cnmfe.m
[^17]: Current demo-parameter documentation. https://caliali-pv.github.io/CaliAli/latest/Functions_doc/CaliAli_demo_parameters/
[^18]: Current getting-started workflow documentation. https://caliali-pv.github.io/CaliAli/latest/Getting_started/