Pith. sign in

REVIEW 4 major objections 7 minor 26 references

CA-Diff: Collaborative Anatomy Diffusion for Brain Tissue Segmentation

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read CA-Diff reports state-of-the-art brain tissue segmentation Dice scores by jointly denoising labels and a registered distance field.

desk verdict Plausible incremental SOTA for brain segmentation, but the consistency loss that drives part of the reported gain is under-specified to the point of being possibly ill-defined; worth peer review with a demand for clarification. read the letter →

arxiv 2506.22882 v1 pith:KBSA2KVI submitted 2025-06-28 eess.IV cs.CVcs.LG

classification eess.IVcs.CVcs.LG
keywords braintissuesegmentationdiffusionmodeldistancefieldanatomicalpriorcollaborativeconsistencylosschannelattentionmagneticresonanceimaging
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that brain-structure segmentation from MRI improves when the diffusion model is given explicit global spatial information about where each voxel sits in the brain. It introduces a distance field as an auxiliary anatomical condition and denoises that field together with the label map, rather than merely feeding the field as an extra input channel. Across three public benchmarks it reports higher Dice and surface-overlap scores than previous CNN, transformer, and diffusion baselines. If the result holds, it suggests that continuous spatial location is a useful conditioning signal for diffusion-based segmentation and that modeling the joint distribution of anatomy and location is better than treating anatomy as a fixed extra input.

What carries the argument

The central object is the anatomical distance field $D_0$: per-voxel coordinate values in a reference brain grid, mapped into each scan by registration, providing global spatial context that patch-based training would otherwise lose. The central mechanism is collaborative diffusion: a single U-Net with two time-embedding streams takes a noisy label and a noisy distance field at two independently sampled timesteps and predicts both original fields, so the model learns the joint distribution $q(D_0, M_0 \mid I)$ rather than only $q(M_0 \mid D_0, I)$. A consistency loss then aligns spatial distance with anatomical similarity by comparing per-voxel predicted label logits, and the time-adapted channel attention module reweights skip-connection features using dynamic convolution whose kernel weights depend on the current diffusion timestep.

What would settle it

Take the trained model and the exact patch sampler; for pairs of patches from different subjects, compute the average anatomical overlap between voxel $n$ in one patch and voxel $n$ in the other after warping both to the reference atlas. If that overlap is low, the consistency loss is pairing non-corresponding locations, and its reported contribution of 0.41 Dice on the first benchmark would need a different explanation.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that collaborative denoising of the segmentation label and an atlas-derived distance field outperforms both standard image-conditioned diffusion and a dual-condition variant that concatenates the distance field with the image. The model learns $\mathbb{E}[D_0, M_0 \mid D_{t_d}, M_{t_m}, I]$ with independently sampled timesteps for the two fields; at inference it sets $t_d = 0$, $t_m = T$, and generates the label with the clean distance field acting as a spatial prior. The paper reports Dice/NSD of 90.63/94.32 on its first benchmark, 89.29/89.85 on the second, and 82.54/78.52 on the third, all above the previous state of the art. The ablations attribute separate gains to the collaborative diffusion scheme, the consistency loss, and the time-adapted channel attention module.

Load-bearing premise

The consistency loss assumes that the same voxel position in two different image patches refers to the same anatomical location, even though the paper does not state that patches from different MRIs are aligned to atlas space before the loss is applied.

Editorial extensions

If this is right

  • An explicit spatial coordinate field can serve as a useful diffusion condition for segmentation because registration supplies it from a reference template rather than requiring the network to infer it from image content alone.
  • Modeling the joint distribution of the auxiliary field and the label helps more than feeding the field as a second input channel, according to the paper's ablation comparison.
  • The reported gains appear with as few as five inference steps, so the method does not require the long sampling chains usually associated with diffusion models.
  • Patch-based training can still carry global spatial information if the distance field is registered into each patch's coordinates.
  • The consistency loss and the time-adaptive attention module each add separate accuracy gains on both evaluated benchmarks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A consequence the paper leaves implicit is that the same collaborative scheme could be applied to other continuous anatomical quantities, such as tissue probability maps or cortical thickness, wherever a stable reference atlas exists.
  • A useful next test would be to replace registered atlas coordinates with a learned deformable correspondence network; if the gains persist, the key ingredient is continuous spatial information rather than the specific registration method.
  • If the consistency loss is computed on native-space patches rather than atlas-aligned patches, the reported gains might reflect a regularization effect rather than anatomical alignment; a direct experiment comparing aligned and unaligned patch pairs would separate these explanations.
  • The paper's own conclusion acknowledges sensitivity to registration precision, which implies that the method's performance ceiling may be partly set by registration quality rather than by the diffusion model itself.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes CA-Diff, a diffusion-based framework for brain MRI segmentation that conditions label generation on a global anatomical distance field. The method introduces a collaborative diffusion process that jointly denoises the label map and the distance field, a spatial-anatomical consistency loss Lsac, and a Time-Adapted Channel Attention module. The authors evaluate on three benchmarks (MALC, SchizBull, Hammers) with three-fold cross-validation and report Dice/NSD improvements over prior state-of-the-art methods, with ablations attributing gains to each component.

Significance. If the reported results hold, CA-Diff would be a new state of the art on three well-known brain-segmentation benchmarks, with a conceptually interesting use of distance fields as a global anatomical condition for diffusion models. The paper's strengths include a fair comparison protocol in which all baselines are rerun under the same three-fold cross-validation, reporting of means and standard deviations in the main comparison (Table I), and a set of ablations for each proposed component. The core concern is that the consistency loss, which is central to the ablations, is not well-defined without an explicit assumption about voxel correspondence across subjects, and several supporting experimental details (error bars in ablations, timestep schedule, inference sampler) are missing.

major comments (4)
  1. [II-D, Eqs. (5)-(7)] The consistency loss Lsac computes per-voxel similarities between two patches i and j with the same voxel index n, and the paper states that these patches 'may be obtained from different MRIs' (Sec. II-D). For SimD to be a 'relative shift distance in the atlas space' and for SimM to compare analogous anatomy, voxel n in both patches must correspond to the same atlas coordinate. The paper, however, only states that the distance field is mapped 'from atlas space to sample space' (Sec. II-B), implying that patches are native-space crops. Without an explicit statement that patches are resampled to a common atlas grid (or that paired patches are extracted from registered volumes), the per-voxel comparisons in Eqs. (5)-(6) are not anatomically aligned, and the BCE target in Eq. (7) is ill-defined. This directly affects the attribution of the +0.57 Dice gain to Lsac in Table II and the cosine-versus-Dice comparison in Table IV. Please clarify whether patches are cropped in atlas space or otherwise aligned before the loss is applied.
  2. [Eq. (7)] The consistency loss applies binary cross-entropy with SimM(i,j) as the target probability. SimM is a cosine similarity of raw network logits, which is not constrained to [0,1] and can be negative; BCE with a negative target is not a valid cross-entropy and its gradient behavior is uncontrolled. The paper should either apply a sigmoid or other normalization to SimM before BCE, or use a regression loss that is well-defined on cosine similarities.
  3. [Tables II-VI] All ablation tables report only point estimates of Dice and NSD, with no standard deviations or significance tests, despite the three-fold cross-validation protocol used in Table I. The claimed component contributions—e.g., the +0.41 Dice from Lsac on MALC and +0.44 Dice from TACA in Table II—are within the scale of fold-to-fold variability visible in Table I (standard deviations of about 0.4 Dice). The ablations should report per-fold results or standard deviations, and ideally significance tests, to demonstrate that the increments are not noise.
  4. [II-C, III-B] There is a train/inference mismatch in the timestep distribution. The text says training samples td and tm uniformly and independently from {0,...,T}, but inference fixes td=0 and uses only the label-denoising branch (Sec. II-C). If T is large, the clean-distance-field condition appears in only 1/(T+1) of training updates, yet at inference it is used 100% of the time. The paper does not state T, does not describe any sampling schedule that up-weights td=0, and does not ablate this choice. This gap affects reproducibility and may explain part of the benefit of the collaborative design, so it should be addressed explicitly.
minor comments (7)
  1. [II-A, Eq. (1)] There is a typo in the sentence following Eq. (1): 'starndard' should be 'standard'. Please fix throughout.
  2. [II-B] The phrase 'equally divide the brain into left and right hemispheres to establish the coordinate system' is unclear; the coordinate system should be described more precisely (e.g., origin at anterior commissure, axes orientation).
  3. [II-C] The claim that the model learns the joint distribution q(D0,M0|I) is overstated, since D0 is a deterministic function of I via registration; this is a conditional multi-task formulation rather than a joint generative model. Please rephrase to avoid overclaiming.
  4. [Table III footnote] The footnote says 'The optimal values differ from those in Table I' but the Collaborative Diffusion row in Table III appears identical to the corresponding row in Table II. Please clarify what is meant by 'incompatible with the ablation setting'.
  5. [III-B] Implementation details omit the value of T, the noise schedule, and the inference sampler (e.g., DDIM vs DDPM). These are needed to reproduce the five-step inference reported in Fig. 4.
  6. [III-B] The sentence 'For 3D tasks, we use a patch size of 96×96×96... while 2D tasks adopt 160×160 patches' is confusing because all three datasets are 3D volumes; either specify which experiments use 2D or remove the 2D sentence.
  7. [Fig. 4] The y-axis of the MALC curve is truncated (starting near 90.15), which visually exaggerates the difference between steps 5 and 10 (90.63 vs 90.65); please use a non-truncated axis or a difference plot.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: CA-Diff's gains come from held-out three-fold cross-validation against external baselines; the distance-field condition is atlas-derived, and the consistency loss is a regularizer rather than a fitted prediction.

full rationale

The paper's central claims are empirical: Table I reports Dice/NSD on held-out subjects under three-fold cross-validation comparing CA-Diff against external methods, and the reported gains are not obtained by fitting a parameter to the test data and renaming it a prediction. The distance field is constructed independently of subject labels: it is built in Colin27 atlas space (anterior commissure origin, left/right hemisphere division) and mapped to each sample by ANTs registration (Sec. II-B, Sec. III-B), so conditioning label generation on it does not smuggle the target labels back in. The collaborative diffusion objective (Eq. 4) and the consistency loss (Eqs. 5-7) are training regularizers over predicted outputs; they do not define the evaluation metric or select the reported numbers by construction. The consistency loss's assumption that paired patches share a common atlas-space voxel grid is not explicitly stated (Sec. II-D says only that patches 'may be obtained from different MRIs'), which is a legitimate correctness/implementation concern that could affect the reliability of the Table II ablation gain, but it is not circularity: the loss is not equivalent to its inputs or to the evaluation protocol. The self-citations in the reference list ([6], [7] for tracking, [8], [9] for diffusion applications) are peripheral and never invoked as load-bearing justification for the anatomy-diffusion design or to forbid alternatives; no author-imported uniqueness theorem is used. The 'joint distribution q(D0, M0|I)' framing overclaims because D0 is deterministic given the image and registration, but that is a conceptual imprecision rather than a circular step. Overall, the derivation chain is self-contained: inputs are atlas coordinates, images, and labels; outputs are segmentation maps tested on held-out subjects against independent baselines.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim relies on the standard diffusion noise model, on the assumption that atlas registration maps distance-field coordinates consistently across subjects, and on the implicit assumption that patch pairs have aligned voxel indices for the consistency loss. The free parameters are design choices (loss weighting, TACA kernel sizes, inference steps) that affect the reported numbers. No new entities are introduced.

free parameters (3)
  • loss weighting = 1.0 for both Ldiff and Lsac
    The total loss L = Ldiff + Lsac uses equal weights without tuning; the relative scale of MSE vs Dice/CE is left as an implicit design choice.
  • TACA kernel sizes = 7, 11, 15
    Kernel sizes for the multi-scale dynamic convolutions are selected by hand and shown in the implementation details.
  • inference steps = 5
    Fig. 4 shows Dice saturates around 5 steps; this value is chosen at inference and affects the reported numbers.
assumptions (3)
  • standard math The diffusion forward process adds Gaussian noise according to Eq. 1, as in standard DDPM.
    Assumed without proof; standard background for all diffusion-based segmentation methods.
  • domain assumption Atlas registration from Colin27 to each subject preserves anatomical correspondence, so the distance field encodes consistent spatial locations across subjects.
    The distance field is constructed in atlas space and mapped to each sample via ANTs registration (Sec. III-B); the method depends on this mapping being accurate.
  • ad hoc to paper Patches from different MRIs have corresponding voxel indices for the consistency loss.
    Eqs. 5-7 sum over voxel index n for patch pairs that may come from different scans; this correspondence is never stated as being enforced by atlas-space resampling.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CA-Diff: Collaborative Anatomy Diffusion for Brain Tissue Segmentation." pith.science (2026). https://pith.science/paper/KBSA2KVI

@misc{pith2026250622882,
  author       = {Pith},
  title        = {Pith review of: CA-Diff: Collaborative Anatomy Diffusion for Brain Tissue Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KBSA2KVI}},
  note         = {Machine review of arXiv:2506.22882}
}
read the original abstract

Segmentation of brain structures from MRI is crucial for evaluating brain morphology, yet existing CNN and transformer-based methods struggle to delineate complex structures accurately. While current diffusion models have shown promise in image segmentation, they are inadequate when applied directly to brain MRI due to neglecting anatomical information. To address this, we propose Collaborative Anatomy Diffusion (CA-Diff), a framework integrating spatial anatomical features to enhance segmentation accuracy of the diffusion model. Specifically, we introduce distance field as an auxiliary anatomical condition to provide global spatial context, alongside a collaborative diffusion process to model its joint distribution with anatomical structures, enabling effective utilization of anatomical features for segmentation. Furthermore, we introduce a consistency loss to refine relationships between the distance field and anatomical structures and design a time adapted channel attention module to enhance the U-Net feature fusion procedure. Extensive experiments show that CA-Diff outperforms state-of-the-art (SOTA) methods.

Figures

Figures reproduced from arXiv: 2506.22882 by the authors.

Figure 1
Figure 1. While Traditional Diffusion method are adapted for MRI label generation solely based on image, the Dual Condition is able to use the anatomical data as additional condition, which is the distance field in our case. We introduce the CA-Diff framework to utilize the Collaborative Diffusion approach to collaboratively learns to denoise both the semantic label map and the distance field to use the anatomical data more e… view at source ↗
Figure 2
Figure 2. The pipeline of our CA-Diff. The distance field and label are perturbed using Gaussian noise, and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative visualizations of the proposed method and baseline models on the MALC dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Ablation of inference step. dition. To leverage this, we develop a collaborative denoising method to link spatial locations with anatomical structures, enhanced by a consistency loss and a time-adapted channel attention module. While CA-Diff demonstrates significant po…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

  1. [1]

    Deep learning for brain mri segmentation: state of the art and future directions,

    Zeynettin Akkus, Alfiia Galimzianova, Assaf Hoogi, Daniel L Rubin, and Bradley J Erickson, “Deep learning for brain mri segmentation: state of the art and future directions,” Journal of digital imaging , vol. 30, pp. 449–459, 2017

  2. [2]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

    Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,” Nature methods, vol. 18, no. 2, pp. 203–211, 2021

  3. [3]

    Unet++: Redesigning skip connections to exploit multiscale features in image segmentation,

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang, “Unet++: Redesigning skip connections to exploit multiscale features in image segmentation,” IEEE transactions on medical imaging, vol. 39, no. 6, pp. 1856–1867, 2019

  4. [4]

    Self-supervised pre-training of swin transformers for 3d medical image analysis,

    Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh, “Self-supervised pre-training of swin transformers for 3d medical image analysis,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 20730–20740

  5. [5]

    Un- etr: Transformers for 3d medical image segmentation,

    Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu, “Un- etr: Transformers for 3d medical image segmentation,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision , 2022, pp. 574–584

  6. [6]

    Trans- former tracking with cyclic shifting window attention,

    Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang, “Trans- former tracking with cyclic shifting window attention,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8791–8800

  7. [7]

    Compact transformer tracker with correlative masked modeling,

    Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang, “Compact transformer tracker with correlative masked modeling,” in Proceedings of the AAAI conference on artificial intelligence , 2023, vol. 37, pp. 2321–2329

  8. [8]

    Diffusiontrack: Diffusion model for multi-object tracking,

    Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang, “Diffusiontrack: Diffusion model for multi-object tracking,” in Proceedings of the AAAI Conference on Artificial Intelligence , 2024, vol. 38, pp. 3991–3999

Show all 26 references
  1. [9]

    Video anomaly detection with motion and appearance guided patch diffusion model,

    Hang Zhou, Jiale Cai, Yuteng Ye, Yonghui Feng, Chenxing Gao, Junqing Yu, Zikai Song, and Wei Yang, “Video anomaly detection with motion and appearance guided patch diffusion model,” arXiv preprint arXiv:2412.09026, 2024

  2. [10]

    Diffbfr: Bootstrapping diffusion model for blind face restoration,

    Xinmin Qiu, Congying Han, ZiCheng Zhang, Bonan Li, Tiande Guo, and Xuecheng Nie, “Diffbfr: Bootstrapping diffusion model for blind face restoration,” in ACM MM, 2023, pp. 7785–7795

  3. [11]

    Diff- unet: A diffusion embedded network for volumetric segmentation,

    Zhaohu Xing, Liang Wan, Huazhu Fu, Guang Yang, and Lei Zhu, “Diff- unet: A diffusion embedded network for volumetric segmentation,”arXiv preprint arXiv:2303.10326, 2023

  4. [12]

    Conditional diffusion model with spatial attention and latent embedding for medical image segmentation,

    Behzad Hejrati, Soumyanil Banerjee, Carri Glide-Hurst, and Ming Dong, “Conditional diffusion model with spatial attention and latent embedding for medical image segmentation,” in International Conference on Med- ical Image Computing and Computer-Assisted Intervention . Springe...

  5. [13]

    Corrdiff: Corrective diffusion model for accurate mri brain tumor segmentation,

    Wenqing Li, Wenhui Huang, and Yuanjie Zheng, “Corrdiff: Corrective diffusion model for accurate mri brain tumor segmentation,” IEEE Journal of Biomedical and Health Informatics , 2024

  6. [14]

    Anatomy-aided deep learning for medical image segmentation: a review,

    Lu Liu, Jelmer M Wolterink, Christoph Brune, and Raymond N J Veld- huis, “Anatomy-aided deep learning for medical image segmentation: a review,” Physics in Medicine & Biology , p. 11TR01, Jun 2021

  7. [15]

    A probabilistic atlas of the human brain: theory and rationale for its development,

    John C Mazziotta, Arthur W Toga, Alan Evans, Peter Fox, Jack Lancaster, et al., “A probabilistic atlas of the human brain: theory and rationale for its development,” Neuroimage, vol. 2, no. 2, pp. 89–101, 1995

  8. [16]

    One transformer fits all distributions in multi-modal diffusion at scale,

    Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu, “One transformer fits all distributions in multi-modal diffusion at scale,” in International Conference on Machine Learning . PMLR, 2023, pp. 1692–1717

  9. [17]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation,

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi, “V-net: Fully convolutional neural networks for volumetric medical image segmentation,” in 2016 fourth international conference on 3D vision (3DV). Ieee, 2016, pp. 565–571

  10. [18]

    Acenet: Anatomical context- encoding network for neuroanatomy segmentation,

    Yuemeng Li, Hongming Li, and Yong Fan, “Acenet: Anatomical context- encoding network for neuroanatomy segmentation,” Medical image analysis, vol. 70, pp. 101991, 2021

  11. [19]

    Can: Context-assisted full attention network for brain tissue segmentation,

    Zhan Li, Chunxia Zhang, Yongqin Zhang, Xiaofeng Wang, Xiaolong Ma, Hai Zhang, and Songdi Wu, “Can: Context-assisted full attention network for brain tissue segmentation,” Medical Image Analysis , vol. 85, pp. 102710, 2023

  12. [20]

    Label-efficient semantic segmentation with diffusion models,

    Dmitry Baranchuk, Ivan Rubachev, Andrey V oynov, Valentin Khrulkov, and Artem Babenko, “Label-efficient semantic segmentation with diffusion models,” arXiv preprint arXiv:2112.03126 , 2021

  13. [21]

    Eca-net: Efficient channel attention for deep convolutional neural networks,

    Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu, “Eca-net: Efficient channel attention for deep convolutional neural networks,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 11534–11542

  14. [22]

    Dynamic convolution: Attention over convolution kernels,

    Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu, “Dynamic convolution: Attention over convolution kernels,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 11030–11039

  15. [23]

    Multi-atlas segmentation with joint label fusion and corrective learning—an open source implementa- tion,

    Hongzhi Wang and Paul A Yushkevich, “Multi-atlas segmentation with joint label fusion and corrective learning—an open source implementa- tion,” Frontiers in neuroinformatics, vol. 7, pp. 27, 2013

  16. [24]

    Candishare: A resource for pediatric neuroimaging data,

    Steven Hodge, Christian Haselgrove, David Kennedy, and Jean Frazier, “Candishare: A resource for pediatric neuroimaging data,” Frontiers in Neuroinformatics, vol. 3, Feb 2010

  17. [25]

    Three-dimensional maximum probability atlas of the human brain, with particular reference to the temporal lobe,

    Alexander Hammers, Richard Allom, Matthias J Koepp, Samantha L Free, Ralph Myers, Louis Lemieux, Tejal N Mitchell, David J Brooks, and John S Duncan, “Three-dimensional maximum probability atlas of the human brain, with particular reference to the temporal lobe,” Human brain m...

  18. [26]

    Enhancement of mr images using registration for signal averaging,

    Colin J Holmes, Rick Hoge, Louis Collins, Roger Woods, Arthur W Toga, and Alan C Evans, “Enhancement of mr images using registration for signal averaging,” Journal of computer assisted tomography, vol. 22, no. 2, pp. 324–333, 1998

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.