SNIF: a Self-Navigator Intelligent Filter correcting off-center and spin-history artifacts in quasi-random radial MRI
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "SNIF: a Self-Navigator Intelligent Filter correcting off-center and spin-history artifacts in quasi-random radial MRI".
Dev: Here's the gist of the paper. SNIF introduces a Self-Navigator Intelligent Filter that models the physical origin of k-space center artifacts to properly estimate and remove them more accurately The artifact affecting the DC-signal SSDDDD generally has two main origins: the jitter of the k-space center (“off-center artifact”) and the cumulative history of the spins (“spin-history artifact”) Both add noise-like artifacts to the signal Modeling Artifact Origins The paper models the DC-signal as a sum of motion-related and artifact components, specifically writing the measured DC-signal as SSDDDD = SSmm + SS00 ⋆ hh The off-center effect can be modeled as a perturbative signal SS00 and the spin-history effect as a convolution kernel hh The spin-history effect is modeled by a convolution kernel h =
hNNRRRRR … h3 h2 h1 h0: , where NNRR is the number of relevant preceding sequence repetitions which influence the current measurement The DC-signal is therefore modeled as SSSDDD = (SSmm + S0) ⋆ h
2: Artifact Estimation Methodology The estimation of the off-center and spin-history artifacts involves a multi-step process using the estimated off-center component SŜ0 Every spoke of a 3D ‘kooshball’ trajectory is characterized by two angles in spherical coordinates: a polar angle,thetatheta, and an azimuthal angle,phiphi (Figure 1a) SSSDDD is processed in Matlab® to extract SŜ0, an estimation of the k-space off-center artifact SS0, with a method based on the C1-constrained interpolation of k-space data from the time domain to a (thetatheta,phiphi) representation of k-space (Figure 1b) Data are filtered by applying successively median, sliding window, and Gaussian low-pass filters before being interpolated back from (thetatheta,phiphi) into the time domain (Figure 1c-d) The estimated convolution kernel h� was then extracted by deconvolving the original DC-signal SSSDDD with SSŜ0 This leads to the estimation of h we want to extract Correction Mechanism The final correction is achieved by subtracting the convolution of the estimated off-center component with the estimated spin-history kernel from the original DC-signal SSSDDD, leading to SSŜmm = SSSDDD − SŜ0 ⋆ h� (Figure 1e) The estimation of h is performed by fitting the general model SSŜ0 ⋆ h� onto the signal SSSDDD in the least squares sense, where SSŜ0 being known and the coefficients of h� being the fitting parameters This process aims to extract more reliable self-navigators by removing both contributions out of the signal Experimental Validation The efficiency of SNIF was evaluated on the real and imaginary DC signal temporal spectra for SSSDDD and SSmm, respectively before and after processing The analysis showed that the DC-signal is clearly improved after the application of SNIF, with a noticeable reduction of the amount of apparent noise enabling a better depiction of the respiratory (long red arrow) and cardiac (short red arrow) motion-related oscillations Peak prominence ratios are enhanced (〈ppr1st〉 > 1 and 〈ppc1st〉 > 1) while the peak widths are reduced (〈wwr1st〉 < 1 and 〈wwc1st〉 < 1), which favors peak selectivity SNIF is expected to be robust and should be applicable to 2D or 3D radial MR acquisition, whatever acquisition parameters are selected, provided quasi-random order of spoke orientations is designed and the condition TT ≫ NNrepTTTT is verified SNIF provides a signal more closely related to the lung and heart motion fields inside the thorax because the sources of biases are physically understood, modelled and filtered with SNIF Ultimately, SNIF is expected to improve the extraction of self-navigators from the DC-signal The second assumption, which states the influence of spin-history, is ascertained by the outcomes of the deconvolution process yielding the estimation h� of h The coefficients of h� at negative indexes (h−1, h−2, …), which could support a backward dependence of SSSDDD on future values of SSŜ0, are experimentally negligible The third assumption considers that the spinhistory artifact follows a linear model based on the convolution kernel h SNIF is thus robust and should be applicable to 2D or 3D radial MR acquisition, whatever acquisition parameters are selected, provided quasi-random order of spoke orientations is designed and the condition TT ≫ NNrepTTTT is verified The paper demonstrates that SNIF is only one to enable the correction of both the offcenter and the spin-history artifacts References
1: J.M.
Rosa: Wild, H.
Dev: Marshall, M.
Rosa: Bock, L.R.
Title and authors: Rosa: So we’re looking at this paper today, "SNIF: a Self-Navigator Intelligent Filter correcting off-center and spin-history artifacts in quasi-random radial MRI." It sounds pretty technical, but basically, they’re tackling two specific kinds of noise that mess up the DC signal in radial MRI.
Dev: Yeah, it’s about those two things: the jitter of the k space center, which they call an off-center artifact, and what they call a spin-history artifact from previous repetitions contaminating the current state. It suggests that we need a better way to estimate these origins to clean up our motion signal.
Taro: From my side, I'm wondering if this is going to work outside of the lab settings where we usually test these systems. If it can handle real-world noise, that’s where it really matters for autonomy.
Rosa: That’s a good point, Taro. The authors are trying to show that by modeling the physical origin of these artifacts—the off-center effect as a perturbative signal and the spin-history as a convolution kernel—they can estimate and remove them more accurately than before.
Dev: They model the DC signal, which they call SSSDDD, as a sum of motion-related and artifact components, specifically writing it as SSSDDD equals SSSmm plus S0 convolved with h > two <ref:2611.13056#pg1>. This tells us that the off-center effect is a perturbative signal called S0 and the spin history effect is modeled by a convolution kernel called h.
Taro: So they are breaking down the signal into these distinct physical parts. That sounds like a solid starting point for understanding how to build something robust.
Rosa: Exactly. The methodology they use involves processing the DC signal in Matlab to extract an estimate of that off-center component, S0, which they call S hat zero, with a method based on C one constrained interpolation of k space data from the time domain to a polar angle theta theta and phi phi representation of k space > one <ref:2611.13056#pg1>.
Dev: They take the data and process it through successive median, sliding window, and Gaussian low-pass filters before interpolating it back from that angular representation into the time domain > one <ref:2611.13056#pg1>. This whole process is designed to get S hat zero.
Taro: So they are essentially mapping the noise onto a coordinate system based on angles and then filtering out what’s not physically coherent in that space. I like that approach for handling quasi-random orientations.
Rosa: Right, and once they have S hat zero, they extract the estimated convolution kernel h by deconvolving the original DC signal with S hat zero > one <ref:2611.13056#pg1>. This gives them an estimate of h which represents that spin history effect we talked about earlier.
The paper's summary: Dev: Now, let’s talk about what they actually found regarding the core problem they were solving in "SNIF: a Self-Navigator Intelligent Filter correcting off-center and spin-history artifacts in quasi-random radial MRI." They are saying that the artifact affecting the DC signal SSSDDD usually has two main origins, and both of these—the off center artifact and the spin history artifact—add noise to the signal.
Rosa: They propose a correction mechanism where they subtract the convolution of their estimated off-center component with their estimated spin-history kernel from the original DC signal SSSDDD to get an improved motion signal called S hat m > one <ref:2611.13056#pg1>.
Taro: That subtraction step is key. It means they aren't just filtering noise; they are actively modeling and subtracting the known physical contamination to leave behind a cleaner measurement of the actual motion.
Dev: They achieve this final correction by fitting their general model, which is S hat zero convolved with h, onto the signal SSSDDD in a least squares sense, where S hat zero is already known and they are just fitting the coefficients of h as parameters > one <ref:2611.13056#pg1>.
Rosa: What’s really interesting from that summary is their expectation. They expect those corrected self navigators to better reflect the underlying motion fields in the patient’s thorax and be less dependent on acquisition biases > two <ref:2611.13056#pg1>.
Taro: So, if it actually makes the motion signal cleaner, that opens up possibilities for retrospective motion correction where we can get a better picture of what happened inside the body.
The paper's improvements: Dev: Let’s look at the specific improvements they claim in "SNIF: a Self-Navigator Intelligent Filter correcting off-center and spin-history artifacts in quasi-random radial MRI." They are showing that this correction helps with the real and imaginary DC signal temporal spectra for SSSDDD and SSmm before and after processing > three <ref:2611.13056#pg1>.
Rosa: The results they show are pretty strong. They report a noticeable reduction in the amount of apparent noise, which allows for a better depiction of both respiratory—the long red arrow—and cardiac—the short red arrow—motion-related oscillations.
Taro: They also claim that the peak prominence ratios are enhanced, meaning alpha p r one is greater than one, and they also reduce the peak widths, meaning alpha w r one is less than one > three <ref:2611.13056#pg1>.
Dev: That reduction in width means it favors peak selectivity when you’re trying to track those heart and lung motions. They are essentially cleaning up the signal so that the motion features are sharper.
Rosa: It seems like they’ve managed to make the respiratory and cardiac self navigators more sensitive to motion, which supports finer MRI motion tracking > three <ref:2611.13056#pg1>. That means better spatial resolution for tracking those physiological movements.
Conclusion: Dev: So, wrapping up on "SNIF: a Self-Navigator Intelligent Filter correcting off-center and spin-history artifacts in quasi-random radial MRI." The paper shows that by understanding the physical origin of the biases, they can model them and filter them out effectively.
Rosa: They achieved this by modeling both the off center effect as a perturbative signal S00 and estimating it through C one constrained interpolation of k space data to angular coordinates theta theta and phi phi > one <ref:2611.13056#pg1>.
Taro: And they did the spin history suppression by deconvolving the DC signal with that estimated off center component to get an estimate of h, which confirms their assumption about causality.
Dev: They confirmed that the coefficients of h at negative indexes, like h minus one or h minus two, are experimentally negligible compared to the positive indexes like h one or two > eight. This supports their assumption that the current magnetization state is influenced by past states but not future ones.
Rosa: And they also made a fourth assumption by introducing the motion component S mm as an additional term in the measured DC signal, justifying it when both off center and spin history artifacts are neglected > seven.
Taro: So, to sum up, SNIF provides a signal that’s more closely related to the lung and heart motion fields inside the thorax because those sources of bias are physically understood, modeled and filtered with SNIF.
Dev: Ultimately, they expect this method to improve the extraction of self navigators from the DC signal > one <ref:2611.13056#pg1>. Provided you design your acquisition parameters with a quasi-random order of spoke orientations and verify that condition T t is greater than N rep t > one <ref:2611.13056#pg1>.
Rosa: It sounds like this is going to be applicable to 2D or three dee radial MR acquisition, no matter what parameters you select, as long as that condition on the quasi-random order of spoke orientations is met <ref:2611.13056#pg1>.
Taro: I just want to say that because they state their limitations, one thing they don't do is not fully capture everything. They flag that the influence of spin history follows a linear model based on the convolution kernel h > one <ref:2611.13056#pg1>.
Dev: That’s fair. So SNIF is robust because it models both types of artifacts and cleans them out, but its robustness relies on those underlying assumptions about linearity being true for the spin history.
Rosa: It’s a solid piece of work that moves us toward more reliable self navigators by focusing on understanding the physics behind the noise in radial MRI.
Tanguy Boucneau, Brice Fernandez, Peder Larson, Luc Darrasse, Xavier Maèitre
Université Paris-Saclay · GE HealthCare · University of California San Francisco
eess.IV
Submitted: 2026-11-09
Updated: 2026-11-09
Comments: 10 pages, 6 figures; supplementary material (93 pages): DC-signal spectra for every reception channel
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 79/100
The gist: SNIF introduces a Self-Navigator Intelligent Filter that models the physical origin of k-space center artifacts to properly estimate and remove them more accurately The artifact affecting the…
Key concepts
- Off-center artifact
- This artifact arises from the jitter of the k-space center during acquisition, modeled as a perturbative signal S00. It is estimated by mapping time-domain k-space data onto angular coordinates (theta, phi) and using C1 interpolation to find its spatial representation.
- Spin-history artifact
- This artifact results from the cumulative history of spins across preceding sequence repetitions, modeled as a convolution kernel h. This kernel is estimated by deconvolving the DC signal with the estimated off-center component, effectively correcting for past measurements.
- DC Signal Modeling
- The measured DC signal is modeled as a sum of motion-related and artifact components: SSSDDD = SSm + S0 ⋆ h. This framework allows SNIF to separate the true motion signal (SSmm) from the noise introduced by off-center jitter and spin history.
- Correction Mechanism
- The final correction subtracts the convolution of the estimated off-center component with the estimated spin-history kernel from the original DC signal. This yields a corrected signal, SSŜm, which is more closely related to lung and heart motion fields.
Terminology
Summary
SNIF introduces a Self-Navigator Intelligent Filter that models the physical origin of k-space center artifacts to properly estimate and remove them more accurately The artifact affecting the DC-signal SSDDDD generally has two main origins: the jitter of the k-space center (“off-center artifact”) and the cumulative history of the spins (“spin-history artifact”) Both add noise-like artifacts to the signal
Modeling Artifact Origins
The paper models the DC-signal as a sum of motion-related and artifact components, specifically writing the measured DC-signal as SSDDDD = SSmm + SS00 ⋆ hh The off-center effect can be modeled as a perturbative signal SS00 and the spin-history effect as a convolution kernel hh The spin-history effect is modeled by a convolution kernel h = [hNNRRRRR … h3 h2 h1 h0], where NNRR is the number of relevant preceding sequence repetitions which influence the current measurement The DC-signal is therefore modeled as SSSDDD = (SSmm + S0) ⋆ h [2]
Artifact Estimation Methodology
The estimation of the off-center and spin-history artifacts involves a multi-step process using the estimated off-center component SŜ0 Every spoke of a 3D ‘kooshball’ trajectory is characterized by two angles in spherical coordinates: a polar angle,thetatheta, and an azimuthal angle,phiphi (Figure 1a) SSSDDD is processed in Matlab® to extract SŜ0, an estimation of the k-space off-center artifact SS0, with a method based on the C1-constrained interpolation of k-space data from the time domain to a (thetatheta,phiphi) representation of k-space (Figure 1b) Data are filtered by applying successively median, sliding window, and Gaussian low-pass filters before being interpolated back from (thetatheta,phiphi) into the time domain (Figure 1c-d) The estimated convolution kernel h� was then extracted by deconvolving the original DC-signal SSSDDD with SSŜ0 This leads to the estimation of h we want to extract
Correction Mechanism
The final correction is achieved by subtracting the convolution of the estimated off-center component with the estimated spin-history kernel from the original DC-signal SSSDDD, leading to SSŜmm = SSSDDD − SŜ0 ⋆ h� (Figure 1e) The estimation of h is performed by fitting the general model SSŜ0 ⋆ h� onto the signal SSSDDD in the least squares sense, where SSŜ0 being known and the coefficients of h� being the fitting parameters This process aims to extract more reliable self-navigators by removing both contributions out of the signal
Experimental Validation
The efficiency of SNIF was evaluated on the real and imaginary DC signal temporal spectra for SSSDDD and SSmm, respectively before and after processing The analysis showed that the DC-signal is clearly improved after the application of SNIF, with a noticeable reduction of the amount of apparent noise enabling a better depiction of the respiratory (long red arrow) and cardiac (short red arrow) motion-related oscillations Peak prominence ratios are enhanced (〈ppr1st〉 > 1 and 〈ppc1st〉 > 1) while the peak widths are reduced (〈wwr1st〉 < 1 and 〈wwc1st〉 < 1), which favors peak selectivity
Conclusion on Robustness
SNIF is expected to be robust and should be applicable to 2D or 3D radial MR acquisition, whatever acquisition parameters are selected, provided quasi-random order of spoke orientations is designed and the condition TT ≫ NNrepTTTT is verified SNIF provides a signal more closely related to the lung and heart motion fields inside the thorax because the sources of biases are physically understood, modelled and filtered with SNIF Ultimately, SNIF is expected to improve the extraction of self-navigators from the DC-signal The second assumption, which states the influence of spin-history, is ascertained by the outcomes of the deconvolution process yielding the estimation h� of h The coefficients of h� at negative indexes (h−1, h−2, …), which could support a backward dependence of SSSDDD on future values of SSŜ0, are experimentally negligible The third assumption considers that the spinhistory artifact follows a linear model based on the convolution kernel h SNIF is thus robust and should be applicable to 2D or 3D radial MR acquisition, whatever acquisition parameters are selected, provided quasi-random order of spoke orientations is designed and the condition TT ≫ NNrepTTTT is verified The paper demonstrates that SNIF is only one to enable the correction of both the offcenter and the spin-history artifacts
References
[1] J.M. Wild, H. Marshall, M. p. 8 Bock, L.R. Schad, P.M. Jakob, M. Puderbach, F. p. 8 Molinari, E. p. 8J.R.V. Beek, J. Biederer
[2] H.-U Kauczor, K. p. 8-F Kreitner
[3] M Zurek, A Bessaad, K Cieslar, Y Crémillieux
[4] P D Gatehouse, G M Bydder
[5] M Garwood
[6] K M Johnson, S B Fain, M L Schiebler, S Nagle
[7] N S Burris, K M Johnson, P E Z Larson
[8] M Tibiletti, A Bianchi, Å Kjørstad, S Wundrak
[9] N S Higano, A D Hahn, J A Tkach
[10] W Jiang, F Ong, K M Johnson, S K Nagle
[11] L Feng, L Axel, H Chandarana, K T Block
[12] R Grimm, S Fürst, I Dregely, C Forman
[13] J Liu, P Spincemaille, N C F Codella
[14] J Delacoste, J Chaptinel, C Beigelman Aubry
[15] B Hiba, N Richard, M Janier, P Croisille
[16] G Buonincontri, C Methner, T Krieg
[17] A Deshmane, M Blaimer, F Breuer, P Jakob
[18] D Piccini, A Littmann, S Nielles‐Vallespin
--- Page 9 ---
The underlying theory of SNIF is built on several strong assumptions The first one is the existence of an off-center component When represented in a (thetatheta,phiphi) coordinate system, the acquired data show a coherent but non-uniform pattern that stands out of what can be named as noise (Figure 1b) On the one hand, this noise can be easily understood as physiologic motions and spin-history effects lose any apparent coherence when represented in (thetatheta,phiphi) because of the quasi-random order of spoke orientations On the other hand, the subject’s MR image is experimentally spatially continuous, so is its spatial Fourier representation Hence, a smooth pattern can be observed and extracted out of the kspace by low-pass filtering the non-coherent components in (thetatheta,phiphi) The second assumption, which states the influence of spin-history, is ascertained by the outcomes of the deconvolution process yielding the estimation h� of h
--- Page 8 ---
The second assumption, which states the influence of spin-history, is ascertained by the outcomes of the deconvolution process yielding the estimation h� of h The coefficients of h� at negative indexes (h−1, h−2, …), which could support a backward dependence of SSSDDD on future values of SSŜ0, are experimentally negligible (Figure 6) with respect to the coefficients of h� at positive indexes (h1, h2, …) Therefore, the current magnetization state is effectively influenced by past magnetization states but not by future ones, in accordance with the causality principle
--- Page 7 ---
Finally, a fourth assumption is made by introducing the motion component SSmm as an additional term in the measured DC-signal It is straightforward to justify when both off-center and spin-history artifacts are neglected In this case, the mean intensity SSSDDD(tt) of an image that would be acquired instantaneously at tt is the sum of i) the mean image intensity without motion 〈S0(tt)〉 = SSS0 and ii) the motion correction term SSmm(tt): SSSDDD(tt) = SSS0 + SSmm(tt) In this situation, the last term characterizes the variation of SSSDDD around SSS0 related to the subject’s motion
--- Page 4 ---
The estimation of the off-center and spin-history artifacts involves a multi-step process using the estimated off-center component SSŜ0 Every spoke of a 3D ‘kooshball’ trajectory is characterized by two angles in spherical coordinates: a polar angle,thetatheta, and an azimuthal angle,phiphi (Figure 1a) SSSDDD is processed in Matlab® to extract SSŜ0, an estimation of the k-space off-center artifact SSS0, with a method based on the C1-constrained interpolation of k-space data from the time domain to a (thetatheta,phiphi) representation of k-space (Figure 1b) Then, data are filtered by applying successively median, sliding window, and Gaussian low-pass filters before being interpolated back from (thetatheta,phiphi) into the time domain (Figure 1c-d) The size of the discretized (thetatheta,phiphi) space in which these filters were applied was 96×192
--- Page 3 ---
Figure 4: Ratios of peak prominences and widths before and after SNIF for the respiratory (top rows) and cardiac (bottom rows) first harmonics mapped onto the 14-channel spine coil array (left) and the 16-channel thorax coil array (right)
--- Page 2 ---
Figure 2 shows a twenty-second close-up of a typical DC-signal extracted from one of the 30 reception channels before and after the correction of off kspace center and spin-history artifacts by SNIF
--- Page 5 ---
Figure 3 shows the imaginary DC-signal associated spectra obtained before (pink) and after SNIF for a. Coil 2, b. Coil 20, and c. p. 1 Coil 28 (see Figure 4 for coil element positioning)
--- Page 6 ---
Figure 5 shows peak frequencies in Hertz of the first three respiratory harmonics mapped onto the 16-channel thorax coil array before and after SNIF
--- Page 4 ---
Figure 4: Ratios of peak prominences and widths before and after SNIF for the respiratory (top rows) and cardiac (bottom rows) first harmonics mapped onto the 14-channel spine coil array (left) and the 16-channel thorax coil array (right)
--- Page 7 ---
Figure 5 shows peak frequencies in Hertz of the first three respiratory harmonics mapped onto the 16-channel thorax coil array before and after SNIF
--- Page 8 ---
Figure 4: Ratios of peak prominences and widths before and after SNIF for the respiratory (top rows) and cardiac (bottom rows) first harmonics mapped onto the 14-channel spine coil array (left) and the 16-channel thorax coil array (right)
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
-
Off-center artifact correction improves respiratory and cardiac selfnavigators by modeling its physical origin as a perturbative signal S00 and estimating it via C1-constrained interpolation of k-space data from the time domain to angular coordinates theta theta and phi phi. This allows for a more accurate extraction of the motion component Smm through subtraction: Smm = SDDDDD - estimated S0 star h.
-
Spin-history artifact suppression is achieved by estimating the convolution kernel h by deconvolving the DC signal with the estimated off-center component, leading to an improved correction of previous sequence repetitions. The system can effectively model this as a linear additive pseudo-noise for rescaled selfnavigators Smm: SDDDDD = Smm + S0 star h.
-
The AI system can extract motion information more sensitively by making respiratory and cardiac selfnavigators more sensitive to motion, which supports finer MRI motion tracking. This is achieved because the method
makes both respiratory and cardiac selfnavigators more sensitive to motion.
Related papers
- HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- FDIR: Fidelity-Anchored Decompressed Image Restoration for Human and Machine Perception
- A Disease-Centric Vision-Language Foundation Model for Precision Oncology in Kidney Cancer
- CADS: A Comprehensive Anatomical Dataset and Segmentation for Whole-Body Anatomy in Computed Tomography
- Task-Aware Bitrate Thresholds for Codec-Robust Ultrasound Segmentation