Fast Slice-Wise Deep Learning Motion Correction for Spinal Cord Functional MRI

arXiv:2611.14677 · cs.CV · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Fast Slice-Wise Deep Learning Motion Correction for Spinal Cord Functional MRI".

Jane: Here's the gist of the paper. Respiratory and bulk motion displace the spinal cord between the volumes of a functional MRI timeseries.

Tom: Our main contribution is a fast, ready-to-use tool, available in SCT v7.4: a lightweight CNN regressing slice-wise in-plane translations.

Jane: This method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0.1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than Spinal Cord Toolbox (SCT).

Tom: The proposed system involves a lightweight convolutional neural network that regresses the in-plane translation aligning each axial slice to a reference.

Title and authors: Tom: So we’re looking at how this whole paper sums up what they did—it’s basically saying they took a lightweight deep learning approach to fix motion in spinal cord fMRI time series, and the main point is that it does it incredibly fast while actually making things better than existing tools.

Jane: It really is just about taking respiratory or bulk motion that messes up the slices between time points and using this specific setup—a convolutional neural network—to predict how far each slice has shifted in the plane of the image.

Tom: Exactly, but what makes it stand out, and I want to get into that because it’s a massive efficiency win. They say they built this model with about ninety-eight thousand parameters in total, which is pretty small for deep learning.

Jane: And the performance numbers are what really sell it; they show improvements on every single metric when you compare it to established tools like SCT and even slice wise AFNI. We’re talking gains of ten to twenty-five percent in signal quality metrics like tSNR, and a reduction in artifacts called DVARS.

Tom: And the most impressive part from a deployment standpoint is the speed; they report it runs about ten seconds per scan, which is thirty times faster than Spinal Cord Toolbox SCT. That’s huge for any real world data processing pipeline.

Jane: They also claim that almost no slice gets displaced by more than one voxel, which is a very tight spatial accuracy constraint for motion correction.

Tom: That precision matters when you’re looking at subtle changes in the spinal cord signal; they even show improvements on outlier slices, cutting them down from fifteen point six percent before correction to just zero point zero six percent after.

Jane: So what this means for a radiologist or a researcher is that they can get cleaner functional data much faster than before, even when dealing with challenging high-field data like the seven Tesla scan mentioned in their results.

Tom: It’s not just about getting a slightly better image; it’s about making the whole workflow viable for higher throughput, which is what Meng cares about when he's thinking about shipping models.

Jane: But they also have to be careful with what they claim; they mention a limitation—since all their methods align to the temporal mean, the metrics are measuring temporal consistency rather than true anatomical accuracy.

Tom: That’s a fair caveat, Jane; you can’t get perfect anatomy without a ground truth for where it *should* be.

Jane: Exactly. The paper is setting a very fast, lightweight baseline for correcting this type of motion, and now we need to see how robust this specific two-stage refinement process holds up when the real world throws in unexpected motion patterns.

The paper's summary: Tom: So, moving past just describing how they built it, we’re looking at what these specific improvements actually mean for the field of spinal cord imaging correction.

Jane: They suggest a two stage inference pipeline, which is pretty neat because it handles both big movements and those smaller residual shifts iteratively to get a final corrected series.

Tom: Yeah, that iterative refinement process is key; Model one handles the large translations using the temporal mean reference, and then Model two refines everything until it gets very close to zero on those small errors.

Jane: And they mention that this setup performs better than slice wise FLIRT in terms of tSNR and DVARS, which is a solid result for signal quality improvement.

Tom: It's not just the speed advantage either; they also note that it keeps the spinal cord within its initial edges for the whole scan, which is much better than what other tools do at times.

Jane: That’s a big deal because when you’re studying functional activity, keeping the anatomy stable across all time points is crucial for getting meaningful results.

Tom: It also performs well on different MRI field strengths; they showed superior performance on the seven Tesla data compared to previous work, which is important since 7T gives you such high resolution <ref:2611.14677#pg1>.

Jane: But we gotta remember that they flag a limitation: because they are aligning everything to the temporal mean, it’s measuring temporal consistency more than perfect spatial accuracy.

Tom: That means if you need absolute anatomical precision down to the voxel level for a specific clinical task, this tool might not be the final answer without some kind of ground truth data.

Jane: It changes things for someone using this tool because it drastically cuts down on the time spent in pre-processing, potentially making high throughput analysis much more feasible.

Tom: And Meng would definitely be interested in that speed; if we can process thousands of subjects a day instead of just a handful, that’s the practical application we need to see.

Jane: So it moves the goalpost for how quickly researchers can get clean spinal cord data from complex acquisition protocols.

Tom: And honestly, looking at what they did here—combining CNN regression with this two-stage refinement—it shows how well these lightweight models can tackle structured motion problems in medical imaging.

Jane: It gives us a clear path forward for developing faster, more robust correction tools that don't have to be as computationally intensive as older registration methods.

Tom: So the next thing we’re going to look at is how they actually trained these models—the synthetic random walk versus the real data training strategy.

The paper's improvements: Tom: So we’ve gone through how they trained this lightweight CNN to predict slice translations, and now we need to wrap up what this whole paper on "Fast Slice-Wise Deep Learning Motion Correction for Spinal Cord Functional MRI" means for us.

Jane: The main takeaway is that this unsupervised approach can correct motion much faster than existing tools while boosting the signal quality metrics like tSNR across different MRI field strengths.

Tom: It really shows how deep learning can provide a viable alternative to more traditional, often slower, registration pipelines when speed and baseline performance are the priorities.

Jane: It’s a practical tool that could significantly reduce the time it takes to process functional spinal cord data in large studies.

Tom: I think if we can get this kind of efficiency into clinical workflows, it opens up a lot more possibilities for longitudinal studies on spinal cord health.

Jane: That’s right. We also have to keep in mind that the authors are clear about the evaluation limits—since they don't have perfect motion ground truth, these improvements are really about temporal consistency rather than absolute spatial accuracy.

Tom: Exactly, so we’re getting a fast tool with great speed and good metrics, but you still need to be mindful of what it doesn't guarantee in terms of anatomical precision.

Jane: And that brings us to the next step, because while this paper focuses on fMRI time series motion, we also have work looking at how misalignment itself can affect model behavior in other areas.

Tom: That’s right, we’re moving from just correcting the data to understanding how misaligned data might actually change what a complex AI model learns.

Jane: We’ll be diving into that next week when we talk about self modeling interventions can modulate emergent misalignment.

Tom: Stick around, because once you see how these models interact with motion, it changes how we think about training AI for real-world data.

Jane: We'll be right back after the break with that discussion on self modeling and emergent misalignment.

Conclusion: Tom: So we’ve covered how they trained this lightweight CNN to predict slice translations, and now we need to wrap up what this paper on "Fast Slice-Wise Deep Learning Motion Correction for Spinal Cord Functional MRI" means for us.

Jane: The main takeaway is that this unsupervised approach can correct motion much faster than existing tools while boosting the signal quality metrics like tSNR across different MRI field strengths.

Tom: It really shows how deep learning can provide a viable alternative to more traditional, often slower, registration pipelines when speed and baseline performance are the priorities.

Jane: It’s a practical tool that could significantly reduce the time it takes to process functional spinal cord data in large studies.

Tom: I think if we can get this kind of efficiency into clinical workflows, it opens up a lot more possibilities for longitudinal studies on spinal cord health.

Jane: That’s right. We also have to keep in mind that the authors are clear about the evaluation limits—since they don't have perfect motion ground truth, these improvements are really about temporal consistency rather than absolute spatial accuracy.

Tom: Exactly, so we’re getting a fast tool with great speed and good metrics, but you still need to be mindful of what it doesn't guarantee in terms of anatomical precision.

Jane: And that brings us to the next step, because while this paper focuses on fMRI time series motion, we also have work looking at how misalignment itself can affect model behavior in other areas.

Tom: That’s right, we’re moving from just correcting the data to understanding how misaligned data might actually change what a complex AI model learns.

Jane: We’ll be diving into that next week when we talk about how self modeling interventions can modulate emergent misalignment.

Tom: Stick around, because once you see how these models interact with motion, it changes how we think about training AI for real-world data.

Jane: We'll be right back after the break with that discussion on self modeling and emergent misalignment.

Lu: The authors set a good baseline for using deep learning to tackle the slice by slice nature of spinal cord acquisition problems.

Meng: For me, the speed and the performance gains on that seven T data are what make this interesting for a startup looking to ship something fast.

Lalam: I’m confident in this direction; the results show a strong correlation between those two-stage models and achieving significant gains over established methods.

Polytechnique Montreal, Montrບl, QC, Canada · Telພcom Paris, Institut Polytechnique de Paris, Palaiseau, France · Faculty of Associated Medical Sciences, Chiang Mai University

cs.CV

Submitted: 2026-11-10

Updated: 2026-11-10

Comments: 4 pages, 3 figures

Code: https://github.com/ivadomed/moco-dl

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 87/100

The gist: Respiratory and bulk motion displace the spinal cord between the volumes of a functional MRI timeseries.

Key concepts

CNN Regressing Slice-Wise Translations
A specialized neural network that learns to predict the exact in-plane shift (translation) of each axial slice relative to a reference slice. This is done by training the network on synthetic and real motion data, allowing it to correct movement that varies across different time points in the MRI scan.
Two-Stage Motion Correction Pipeline
A multi-step correction process involving two models: M1 handles large, persistent movements across the entire series, and M2 refines small residual errors iteratively. This ensures a robust correction by first addressing major shifts before focusing on subtle misalignments.
NCC Loss (Normalized Cross-Correlation)
A mathematical measure used during training to quantify how well the predicted slice aligns with the reference image. The loss function guides the CNN to minimize this correlation error, ensuring that the corrected slices look as similar as possible to a perfectly aligned reference.
tSNR and DVARS Improvement
These metrics measure signal quality and variability in functional MRI data. The proposed method consistently improves tSNR (signal-to-noise ratio) and reduces DVARS (temporal variability), indicating that the motion correction leads to cleaner, more reliable functional data.

Terminology

Summary

Respiratory and bulk motion displace the spinal cord between the volumes of a functional MRI timeseries. p. 0

Where the paper says it

FAST SLICE-WISE DEEP LEARNING MOTION CORRECTION FOR SPINAL CORD FUNCTIONAL MRI Thomas Dagonneau1,2 , Sandrine Bedard ´ 1 , Nontharat Tucksinapinunchai3 , Uten Yarach3 , Lo¨ıc Le Folgoc2 , Pietro Gori2 , Julien Cohen-Adad1 1Polytechnique Montreal, Montr ´ eal, QC, Canada ´ 2Tel´ ecom Paris, Institut Polytechnique de Paris, Palaiseau, France ´ 3Faculty of Associated Medical Sciences, Chiang Mai University, Chiang Mai, Thailand ABSTRACT Respiratory and bulk motion displace the spinal cord between the volumes of a functional MRI timeseries. Current methods take several minutes per time series, and general-purpose brain tools register whole volumes, whereas the cord moves mainly slice by slice.

Our main contribution is a fast, ready-to-use tool, available in SCT v7. p. 0
Where the paper says it

DeepRetroMoCo [8], an unsupervised network predicting dense deformation fields, was trained and tested on a single 3T dataset and is released as research code rather than a ready-to-use tool. Our main contribution is a fast, ready-to-use tool, available in SCT v7.

4: a lightweight CNN regressing slice-wise in-plane translations. p. 0
Where the paper says it

Our main contribution is a fast, ready-to-use tool, available in SCT v7. 4: a lightweight CNN regressing slice-wise in-plane translations, trained without labels on multi-site 3T and 7T data with random-walk motion augmentation, and applied in two stages (large, then residual motion) with an NCC check that rejects harmful predictions.

This method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0. p. 0
Where the paper says it

Two models, trained without motion labels, correct large and then residual motion. On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0.

1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than Spinal Cord Toolbox (SCT). p. 0
Where the paper says it

On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0. 1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than Spinal Cord Toolbox (SCT).

Methodology of the Correction Tool

The proposed system involves a lightweight convolutional neural network that regresses the in-plane translation aligning each axial slice to a reference. The architecture consists of four 3×3 stride-2 convolutions with LeakyReLU (16 to 128 channels), adaptive average pooling and a linear layer for about 98k parameters in total. p. 0

Where the paper says it

Network and preprocessing For each axial slice of each volume, the network takes the slice to correct, It, and a reference slice, R (by default the temporal mean), stacked as channels, and outputs the translation (tx, ty) aligning It to R (Fig. The architecture consists of four 3×3 stride-2 convolutions with LeakyReLU (16 to 128 channels), adaptive average pooling and a linear layer for about 98k parameters in total.

Images are first cropped around the cord with sc-crop (detection on the first volume, same box for all volumes), which removes background structures that move differently from the cord and speeds up inference. p. 0
Where the paper says it

The architecture consists of four 3×3 stride-2 convolutions with LeakyReLU (16 to 128 channels), adaptive average pooling and a linear layer for about 98k parameters in total. Images are first cropped around the cord with sc-crop (detection on the first volume, same box for all volumes), which removes background structures that move differently from the cord and speeds up inference.

Each slice is padded or cropped to 64×64 and its intensities are clipped to the 1st–99th percentiles and rescaled to [0, 1]. p. 0
Where the paper says it

Images are first cropped around the cord with sc-crop (detection on the first volume, same box for all volumes), which removes background structures that move differently from the cord and speeds up inference. Each slice is padded or cropped to 64×64 and its intensities are clipped to the 1st–99th percentiles and rescaled to [0, 1].

Training Strategy for the Deep Learning Model

Two models share this architecture: M1 for large, persistent movements and M2 for small residual ones. For M1, a synthetic random-walk translation is added to each series where tk = tk−1 + εk, εk ∼ N (0, σ2), independently along x and y, with σ = 3 voxels and each component clipped to [−3, 3] voxels. p. 1

Where the paper says it

of runs was 420, which were split by subject into 337 for training and 83 for validation, giving 2,603,976 and 683,041 slice pairs. For M1, a synthetic random-walk translation is added to each series: tk = tk−1 + εk, εk ∼ N (0, σ2 ), independently along x and y, with σ = 3 voxels and each component clipped to [−3, 3] voxels, so that displacements drift and persist as real motion does.

M2 is trained on the real series without synthetic motion to focus on small displacements present in the data. p. 1
Where the paper says it

For M1, a synthetic random-walk translation is added to each series: tk = tk−1 + εk, εk ∼ N (0, σ2 ), independently along x and y, with σ = 3 voxels and each component clipped to [−3, 3] voxels, so that displacements drift and persist as real motion does. M2 is trained on the real series without synthetic motion, to focus on the small displacements present in the data.

The loss for M1, L1 = Lg + Ll, combines a global NCC over the non-zero pixels (Lg) with a local NCC (5×5 window [9], Ll) sensitive to subvoxel misalignment. p. 1
Where the paper says it

The loss compares the warped slice with the mean R of the series after data augmentation. For M1, the loss L1 = Lg + Ll combines a global NCC over the non-zero pixels, Lg, which keeps a useful gradient for large displacements, with a local NCC (5×5 window [9]), Ll , sensitive to subvoxel misalignment.

Performance Gains on Spinal Cord Data

The method demonstrates significant improvements across multiple metrics when compared to existing tools like sct fmri moco and slice-wise AFNI. On the 3T dataset (ds008784), the proposed method improves on sct fmri moco on every metric (all p < 10−12) and is 24 times faster. p. 0

Where the paper says it

On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0. 1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than Spinal Cord Toolbox (SCT).

For the 7T data (ds007067), our method improves on sct fmri moco on every metric (tSNR 12. p. 2
Where the paper says it

Results 3T data. With bilinear interpolation (Table 1), our method improves on sct fmri moco on every metric (all p < 10−12) and is 24 times faster.

21 vs. 9.76, DVARS 12. p. 0
Where the paper says it

Two models, trained without motion labels, correct large and then residual motion. On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0.

40 vs. 20.17, all p < 10−9) and is 38 times faster. p. 0
Where the paper says it

On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0. 1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than Spinal Cord Toolbox (SCT).

Limitations and Future Directions

The main limitation identified is the evaluation, as with no ground truth, and all methods aligning to the temporal mean, the metrics measure temporal consistency rather than anatomical accuracy. Furthermore, the network estimates translations only whereas FLIRT also corrects in-plane rotation which may explain its slightly sharper images at 3T. p. 3

Where the paper says it

Simulated motion of known amplitude would quantify accuracy directly. Also, the network estimates translations only, whereas FLIRT also corrects in-plane rotation, which may explain its slightly sharper images at 3T.

Future work will further validate in other databases and extend the method to spinal cord diffusion MRI. p. 3
Where the paper says it

Also, the network estimates translations only, whereas FLIRT also corrects in-plane rotation, which may explain its slightly sharper images at 3T. Future work will further validate in other databases, and extend the method to spinal cord diffusion MRI.

Conclusion

Compared with sct fmri moco, a small unsupervised CNN with 98k parameters corrects spinal cord motion in fMRI timeseries better and runs 24 to 38 times faster. It performs on par with AFNI and better than slice-wise FLIRT, although more extensive tests are warranted. p. 3

Where the paper says it

DISCUSSION AND CONCLUSION Compared with sct fmri moco, a small unsupervised CNN with 98k parameters corrects spinal cord motion in fMRI timeseries better and runs 24 to 38 times faster. It performs on par with AFNI and better than slice-wise FLIRT, although more extensive tests are warranted.

Acknowledgments

This work was funded by AMX (France), CIHR, NSERC, Praxis and NVIDIA. The authors have no relevant financial or non-financial interests to disclose. p. 3

Where the paper says it

ACKNOWLEDGMENTS This work was funded by AMX (France), CIHR, NSERC, Praxis and NVIDIA. The authors have no relevant financial or non-financial interests to disclose.

References

[1] M. Fratini, M. p. 3

Where the paper says it

REFERENCES [1] M. Fratini, M.

Moraschi, B. p. 3
Where the paper says it

Fratini, M. Moraschi, B.

Maraviglia, and F. p. 3
Where the paper says it

Moraschi, B. Maraviglia, and F.

Giove, “On the impact of physiological noise in spinal cord functional MRI,” J. p. 3
Where the paper says it

Maraviglia, and F. Giove, “On the impact of physiological noise in spinal cord functional MRI,” J.

Magn. Reson. Imaging, vol 40, no 4, pp 770–777. p. 3
Where the paper says it

Imaging, vol. 770–777, Oct.

[2] J. Cohen-Adad et al. p. 3

Where the paper says it

770–777, Oct. Cohen-Adad et al.

, “Slice-by-slice motion correction in spinal cord fMRI: SliceCorr,” in Proc. p. 3
Where the paper says it

Cohen-Adad et al. , “Slice-by-slice motion correction in spinal cord fMRI: SliceCorr,” in Proc.

Intl. Soc. Mag. Reson. Med (ISMRM), 2009, vol 17. p. 3
Where the paper says it

, “Slice-by-slice motion correction in spinal cord fMRI: SliceCorr,” in Proc. (ISMRM), 2009, vol.

[3] B. De Leener et al. p. 3

Where the paper says it

(ISMRM), 2009, vol. De Leener et al.

, “SCT: Spinal Cord Toolbox, an open-source software for processing spinal cord MRI data,” NeuroImage, vol 145, Part A, pp 24–43, 2017. p. 3
Where the paper says it

, “SCT: Spinal Cord Toolbox, an open-source software for processing spinal cord MRI data,” NeuroImage, vol. 145, Part A, pp.

[4] B. B. Avants, N. J. Tustison, G. p. 3

Where the paper says it

24–43, 2017. Tustison, G.

Song, P. A. Cook, A. Klein, and J. p. 3
Where the paper says it

Tustison, G. Klein, and J.

C. Gee, “A reproducible evaluation of ANTs similarity metric performance in brain image registration,” NeuroImage, vol 54, no 3, pp 2033–2044. p. 3
Where the paper says it

Gee, “A reproducible evaluation of ANTs similarity metric performance in brain image registration,” NeuroImage, vol. 2033–2044, 2011.

[5] R. W. Cox, “AFNI: software for analysis and visualization of functional magnetic resonance neuroimages,” Comput. p. 3

Where the paper says it

2033–2044, 2011. Cox, “AFNI: software for analysis and visualization of functional magnetic resonance neuroimages,” Comput.

Biomed. Res., vol 29, no 3, pp 162–173, 1996. p. 3
Where the paper says it

Cox, “AFNI: software for analysis and visualization of functional magnetic resonance neuroimages,” Comput. 162–173, 1996.

[6] M. Jenkinson, P. p. 3

Where the paper says it

162–173, 1996. Jenkinson, P.

Bannister, M. p. 3
Where the paper says it

Jenkinson, P. Bannister, M.

Brady, and S. p. 3
Where the paper says it

Bannister, M. Brady, and S.

Smith, “Improved optimization for the robust and accurate linear registration and motion correction of brain images,” NeuroImage, vol 17, no 2, pp 825–841, 2002. p. 3
Where the paper says it

Smith, “Improved optimization for the robust and accurate linear registration and motion correction of brain images,” NeuroImage, vol. 825–841, 2002.

[7] D. Rangaprakash and R. p. 3

Where the paper says it

825–841, 2002. Rangaprakash and R.

L. Barry, “Neptune: a toolbox for spinal cord functional MRI data processing and quality assurance,” in Proc. p. 3
Where the paper says it

Rangaprakash and R. Barry, “Neptune: a toolbox for spinal cord functional MRI data processing and quality assurance,” in Proc.

Intl. Soc. Mag. Reson Med (ISMRM), 2022, p 0396. p. 3
Where the paper says it

Barry, “Neptune: a toolbox for spinal cord functional MRI data processing and quality assurance,” in Proc. (ISMRM), 2022, p.

[8] M. Mobarak-Abadi et al. p. 3

Where the paper says it

(ISMRM), 2022, p. Mobarak-Abadi et al.

, “DeepRetroMoCo: deep neural network-based retrospective motion correction algorithm for spinal cord fMRI,” Front. p. 3
Where the paper says it

Mobarak-Abadi et al. , “DeepRetroMoCo: deep neural network-based retrospective motion correction algorithm for spinal cord functional MRI,” Front.

Psychiatry, vol 15, pp 1323109, June 2024. p. 3
Where the paper says it

Psychiatry, vol. 1323109, June 2024.

[9] M. J. Cardoso et al. p. 3

Where the paper says it

1323109, June 2024. Cardoso et al.

, “MONAI: An open-source framework for deep learning in healthcare,” arXiv preprint arXiv:2211. p. 3
Where the paper says it

Cardoso et al. , “MONAI: An open-source framework for deep learning in healthcare,” arXiv preprint arXiv:2211.

02701, 2022. p. 3
Where the paper says it

, “MONAI: An open-source framework for deep learning in healthcare,” arXiv preprint arXiv:2211. 02701, 2022.

[10] S. Bedard et al. p. 0

Where the paper says it

FAST SLICE-WISE DEEP LEARNING MOTION CORRECTION FOR SPINAL CORD FUNCTIONAL MRI Thomas Dagonneau1,2 , Sandrine Bedard ´ 1 , Nontharat Tucksinapinunchai3 , Uten Yarach3 , Lo¨ıc Le Folgoc2 , Pietro Gori2 , Julien Cohen-Adad1 1Polytechnique Montreal, Montr ´ eal, QC, Canada ´ 2Tel´ ecom Paris, Institut Polytechnique de Paris, Palaiseau, France ´ 3Faculty of Associated Medical Sciences, Chiang Mai University, Chiang Mai, Thailand ABSTRACT Respiratory and bulk motion displace the spinal cord between the volumes of a functional MRI timeseries. Current methods take several minutes per time series, and general-purpose brain tools register whole volumes, whereas the cord moves mainly slice by slice.

, “Spatial distribution of spinal cord fMRI activity with electrocutaneous stimulation,” Imaging Neuroscience, vol 4, pp IMAG. p. 3
Where the paper says it

Bedard et al. , “Spatial distribution of spinal cord ´ fMRI activity with electrocutaneous stimulation,” Imaging Neuroscience, vol.

a.1383, 2026. p. 3
Where the paper says it

, “Spatial distribution of spinal cord ´ fMRI activity with electrocutaneous stimulation,” Imaging Neuroscience, vol. 1383, 2026.

[11] A. D’Astous et al. p. 3

Where the paper says it

1383, 2026. D’Astous et al.

, “Spine fMRI 7T,” OpenNeuro, doi:10. p. 3
Where the paper says it

D’Astous et al. , “Spine fMRI 7T,” OpenNeuro, doi:10.

18112/openneuro. p. 3
Where the paper says it

, “Spine fMRI 7T,” OpenNeuro, doi:10. 18112/openneuro.

ds007067.v4.0.0, 2026.

[N] The paper is 2611. p. 0

Where the paper says it

The method is tested on two other public datasets, unseen during training (ds008784 and ds007067, Section 3. Some of the datasets included multiple runs per subject: the total number arXiv:2611.

--- Page 5 ---

(a) Training raw fMRI t = 1…T Crop + normalize Random-walk motion tk = tk−1 + εk t M1: random walk M2: no synthetic motion tx ty augmented series mean R slice I t 2×H×W CNN (tx, ty) Warp NCC loss M1: global + local M2: local backprop (b) Two-stage inference raw fMRI t = 1…T Crop + normalize I t, R Model 1 large motion Warp NCC check Model 2 small motion Warp NCC check × N, update R corrected fMRI Conv 3×3/s2 + LeakyReLU (×4) Adaptive avg pool FC 128→2 reference path transforms composed, applied once to raw data Fig. 1 Overview. (a) Training: the CNN predicts (tx, ty) from a slice It and the temporal mean R; an NCC loss compares the warped slice with R. p. 1

Where the paper says it

(a) Training raw fMRI t = 1…T Crop + normalize Random-walk motion tk = tk−1 + εk t M1: random walk M2: no synthetic motion tx ty augmented series mean R slice I t 2×H×W CNN (tx , ty ) Warp NCC loss M1: global + local M2: local backprop (b) Two-stage inference raw fMRI t = 1…T Crop + normalize I t , R Model 1 large motion Warp NCC check Model 2 small motion Warp NCC check × N, update R corrected fMRI Conv 3×3/s2 + LeakyReLU (×4) Adaptive avg pool FC 128→2 reference path transforms composed, applied once to raw Fig. (a) Training: the CNN predicts (tx, ty) from a slice It and the temporal mean R; an NCC loss compares the warped slice with R.

--- Page 6 ---

(b) Two-stage inference raw fMRI t = 1…T Crop + normalize I t, R Model 1 large motion Warp NCC check Model 2 small motion Warp NCC check × N, update R corrected fMRI Conv 3×3/s2 + LeakyReLU (×4) Adaptive avg pool FC 128→2 reference path transforms composed, applied once to raw data Fig. 1 Overview. (b) Two-stage inference M1 predicts a translation for every slice, with R the temporal mean; any translation that lowers the NCC between the slice and R is replaced by the identity. p. 1

Where the paper says it

Both models are trained with Adam (learning rate 10−4 , 100 epochs, batch size 128), keeping the checkpoint with the lowest validation loss on real pairs; each takes about 13 h on an NVIDIA RTX PRO 6000 GPU. Two-stage iterative inference M1 predicts a translation for every slice, with R the temporal mean; any translation that lowers the NCC between the slice and R is replaced by the identity.

M2 then refines every slice iteratively: R is recomputed from the corrected series, and Sct AFNI FLIRT Ours M2 predicts a residual translation, filtered by the same check and composed with the current transform. p. 1
Where the paper says it

Two-stage iterative inference M1 predicts a translation for every slice, with R the temporal mean; any translation that lowers the NCC between the slice and R is replaced by the identity. M2 then refines every slice iteratively: R is recomputed from the corrected series, and SCT AFNI FLIRT Ours −50 −25 0 25 50 75 100 tSNR change vs.

p. 1
Where the paper says it

Two-stage iterative inference M1 predicts a translation for every slice, with R the temporal mean; any translation that lowers the NCC between the slice and R is replaced by the identity. M2 then refines every slice iteratively: R is recomputed from the corrected series, and SCT AFNI FLIRT Ours −50 −25 0 25 50 75 100 tSNR change vs.

Refinement stops after N = 6 iterations, when the mean parameter change falls below 10−3, or when it increases (the previous transform is then kept). p. 1
Where the paper says it

M2 predicts a residual translation, filtered by the same check and composed with the current transform. Refinement stops after N = 6 iterations, when the mean parameter change falls below 10−3 , or when it increases (the previous transform is then kept).

p. 1
Where the paper says it

M2 predicts a residual translation, filtered by the same check and composed with the current transform. Refinement stops after N = 6 iterations, when the mean parameter change falls below 10−3 , or when it increases (the previous transform is then kept).

The composed transform is applied once to the raw, uncropped slices. p. 1
Where the paper says it

Refinement stops after N = 6 iterations, when the mean parameter change falls below 10−3 , or when it increases (the previous transform is then kept). The composed transform is applied once to the raw, uncropped slices.

p. 1
Where the paper says it

Refinement stops after N = 6 iterations, when the mean parameter change falls below 10−3 , or when it increases (the previous transform is then kept). The composed transform is applied once to the raw, uncropped slices.

How the training signal is built

The training process involves augmenting the series with random-walk motion for M1 where tk = tk−1 + εk, εk ∼ N (0, σ2), independently along x and y, with σ = 3 voxels and each component clipped to [−3, 3] voxels. M2 is trained on the real series without synthetic motion. p. 1

Where the paper says it

For M1, a synthetic random-walk translation is added to each series: tk = tk−1 + εk, εk ∼ N (0, σ2 ), independently along x and y, with σ = 3 voxels and each component clipped to [−3, 3] voxels, so that displacements drift and persist as real motion does. M2 is trained on the real series without synthetic motion, to focus on the small displacements present in the data.

The loss for M2 compares the warped slice with the mean R of the series after data augmentation. p. 1
Where the paper says it

M2 is trained on the real series without synthetic motion, to focus on the small displacements present in the data. The loss compares the warped slice with the mean R of the series after data augmentation.

Where the gains come from

The two-stage inference involves Model 1 predicting a translation for every slice, with R the temporal mean; any translation that lowers the NCC between the slice and R is replaced by the identity. M2 then refines every slice iteratively: R is recomputed from the corrected series, and Sct AFNI FLIRT Ours M2 predicts a residual translation, filtered by the same check and composed with the current transform. Refinement stops after N = 6 iterations, when the mean parameter change falls below 10−3, or when it increases (the previous transform is then kept). The composed transform is applied once to the raw, uncropped slices.

Final Summary

The proposed method improves on sct fmri moco on every metric (all p < 10−12) and is 24 times faster. It improves tSNR in every scan, whereas sct fmri moco lowers it in some (Fig. p. 2

Where the paper says it

With bilinear interpolation (Table 1), our method improves on sct fmri moco on every metric (all p < 10−12) and is 24 times faster. It improves tSNR in every scan, whereas sct fmri moco lowers it in some (Fig.

2). It improves on AFNI on every metric but sharpness (p < 10−7) and is 25 times faster. p. 0
Where the paper says it

On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0. 1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than Spinal Cord Toolbox (SCT).

It also improves on FLIRT in tSNR, DVARS and outlier slices (p < 0. p. 2
Where the paper says it

It improves on AFNI on every metric but sharpness (p < 10−7 ). It also improves on FLIRT in tSNR, DVARS and outlier slices (p < 0.

02) and is 25 times faster. p. 0
Where the paper says it

On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0. 1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than Spinal Cord Toolbox (SCT).

--- Page 3 ---

Raw Mean Profile along the dashed line, over time Volume tSNR SCT AFNI FLIRT Ours -50 -25 0 25 50 75 100 120.

--- Page 4 ---

Table 2. Results on ds007067 (7T, ds008784), as in Table 2 Method tSNR↑ DVARS↓ Sharp. p. 3

Where the paper says it

Results on ds007067 (7T, 59 scans), as in Table 1. Method tSNR↑ DVARS↓ Sharp.

↑ Outl.↓ Time (s) Raw 6. p. 2
Where the paper says it

Raw Mean Profile along the dashed line, over time 3. 5 tSNR SCT 3.

87∗ 26.31∗ 0.170∗ 15.62∗ – sct fmri moco 9. p. 0
Where the paper says it

The Spinal Cord Toolbox (SCT, v7. 3) [3] provides a slicewise correction, sct fmri moco, built on ANTs [4], which takes 4–7 min per time series.

76∗ 20.17∗ 0.170 star. AFNI 12.09 12.23 0.196 star. FSL FLIRT 11. p. 0
Where the paper says it

On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS −20 to −38%), leaves almost no slice displaced by more than one voxel (< 0. 1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than Spinal Cord Toolbox (SCT).

78 star. Ours 5 times faster. p. 3
Where the paper says it

1∗ Ours 12. 7 times faster.

--- Page 5 ---

Fig. 3 Example scan with severe motion, one slice (40% of slices displaced) in which the cord jumps by several voxels at volume 37 Only our method keeps the spinal cord within its initial edges for the whole scan: FLIRT reduces the jump but leaves the cord shifted by about one voxel afterwards, and sct fmri moco and AFNI leave it displaced. p. 3

Where the paper says it

3 shows an example scan (40% of slices displaced) in which the cord jumps by several voxels at volume 37. Only our method keeps the spinal cord within its initial edges for the whole scan: FLIRT reduces the jump but leaves the cord shifted by about one voxel afterwards, and sct fmri moco and AFNI leave it displaced.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Lightweight CNN for Slice-wise Translation Regression: The proposed network regresses in-plane translations aligning each axial slice to a reference, achieving a lightweight convolutional neural network that regresses the in-plane translation aligning each axial slice to a reference. This allows for correction of motion that varies along the length of the cord, which is acquired at different times.

  2. Two-Stage Motion Correction Pipeline: The system employs two models, M1 for large, persistent movements and M2 for small residual ones, performing two stages (large, then residual motion). This iterative refinement ensures that translations that lower the NCC between a slice and the temporal mean are discarded to produce a final corrected fMRI time series.

  3. Enhanced Performance on Different Field Strengths: The method shows superior performance across different MRI strengths, improving on sct fmri moco on every metric for 3T data and significantly better for 7T data, such as achieving tSNR 12.21 vs. 9.76 and reducing outlier slices from 15.6% before correction to 0.06% in the ds007067 dataset at 7T.

  4. Increased Computational Efficiency: The proposed method is significantly faster, running about 10 s per scan, about 30 times faster than Spinal Cord Toolbox (SCT) compared to existing tools, making it viable for high-throughput processing of spinal cord fMRI time series.

Abstract

Respiratory and bulk motion displace the spinal cord between the volumes of a functional MRI time series. Current methods such as the Spinal Cord Toolbox (SCT) take several minutes per time series, and general-purpose brain tools register whole volumes, whereas the cord moves mainly slice by slice. We propose a lightweight convolutional neural network that regresses the in-plane translation aligning each axial slice to a reference. Two models, trained without motion labels, correct large and then residual motion. On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS-20 to-38%), leaves almost no slice displaced by more than one voxel (<0.1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than SCT. The code is available at https://github.com/ivadomed/moco-dl and the method is available in SCT v7.4.

Sources

Related papers