pith. sign in

arxiv: 1907.00068 · v1 · pith:Y6WLNZKBnew · submitted 2019-06-28 · 💻 cs.CV · eess.IV

On Reducing Negative Jacobian Determinant of the Deformation Predicted by Deep Registration Networks

Dongyang Kuang This is my paper

Pith reviewed 2026-05-25 13:23 UTC · model grok-4.3

classification 💻 cs.CV eess.IV
keywords image registrationdeep learningJacobian determinantdeformation fieldsdiffeomorphismmedical imagingunsupervised learningfolding
0
0 comments X

The pith

Two training mechanisms reduce folding in deformations from deep registration networks without architecture or hyperparameter changes.

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

Medical image registration seeks transformations that align images while remaining smooth and invertible. Deep unsupervised networks often violate invertibility by producing folds, shown by negative Jacobian determinants at multiple voxels. Existing fixes add explicit penalties that demand careful loss weighting. The paper tests two separate training mechanisms that can be added on top of any baseline registration loss. These mechanisms cut the count of folding locations on standard medical datasets while leaving the network and its hyperparameters untouched.

Core claim

The paper shows that two independent training mechanisms, introduced without altering the baseline loss terms, hyperparameters, or network architecture, substantially lower the number of voxels where the Jacobian determinant of the predicted deformation is negative.

What carries the argument

Two independent training mechanisms that encourage avoidance of negative Jacobian determinants during optimization of the registration network.

If this is right

  • Fewer voxels exhibit non-invertible folding in the output deformation fields.
  • The mechanisms integrate directly with existing unsupervised registration losses.
  • No retuning of regularization weights or network design is required.
  • Smoother deformations are obtained on typical medical image pairs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same mechanisms could be tested on non-medical registration problems such as satellite or microscopy alignment.
  • They might permit weaker spatial smoothness penalties while still controlling folds.
  • Post-processing steps that repair folds could become less necessary.

Load-bearing premise

The mechanisms can be added independently of existing losses and will reduce negative Jacobians on medical datasets without creating new failure modes or requiring per-dataset tuning.

What would settle it

Running the baseline registration network with either mechanism on a representative medical dataset and finding no measurable drop in the count of negative Jacobian locations.

Figures

Figures reproduced from arXiv: 1907.00068 by Dongyang Kuang.

Figure 1
Figure 1. Figure 1: An overview of the registration network usually used for registration. The popu￾lar U-net architecture[9] is used as the deformation unit for generating the displacement field. These work emphasize more on the accuracy and efficiency of registration when compared to classical methods but usually did not put equal emphasis on checking geometric properties such as smoothness, invertibility or orientation pre… view at source ↗
Figure 2
Figure 2. Figure 2: A snapshot of at the same location of the projected warped grid with different regularization strength. From left to right, the network is trained with λ = 1, 2, 4 separately. tive Jacobian determinant issue of unsupervised registration networks. The two ideas inspect the problem from a different angle by altering the training mech￾anism instead of tuning hyper-parameters or changing the commonly adopted l… view at source ↗
Figure 3
Figure 3. Figure 3: A diagram illustrating the cycle consistent design. While it is straightforward that this design directly addresses the invertibil￾ity of the network, the cycle constraint also contributes to the task of learning a smooth solution in an indirect way: the design regularizes the network by forcing the spatial transformer to learn a solu￾tion and its inverse at the same time. This helps the network rule out p… view at source ↗
Figure 4
Figure 4. Figure 4: A diagram showing the activity of the network during training phases. Col￾ored part in the above figure is the base￾line network, greyed and dashed part is the attachable refinement network. D: the network producing displacement field. S: the sampling module. R: the network for learning possible corrections needed for smoother field with less “folding”. Training is happening alternatively between these two… view at source ↗
Figure 5
Figure 5. Figure 5: Left: Sample brain volume and 2 labels. Right: Mean dice scores of different methods on selected regions. Each point is the mean dice score averaged over corre￾sponding ROI labels per registration pair instead of over the union of labels in that region. Results from SyNQuick algorithm in the ANTs package are also listed for better interpreting these dice scores, but not for the purpose of comparison. Metho… view at source ↗
Figure 6
Figure 6. Figure 6: Determinant of Jacobian map and the warped grid projected on the same slice. From left to right: the baseline VoxelMorph prediction, baseline with cycle consistent design and baseline with refinement module. Locations where the determinants are negative are shown in red. words, more than 90% of the unsatisfactory locations happening in the baseline prediction are eliminated. With the refinement module, tho… view at source ↗
read the original abstract

Image registration is a fundamental step in medical image analysis. Ideally, the transformation that registers one image to another should be a diffeomorphism that is both invertible and smooth. Traditional methods like geodesic shooting approach the problem via differential geometry, with theoretical guarantees that the resulting transformation will be smooth and invertible. Most previous research using unsupervised deep neural networks for registration have used a local smoothness constraint (typically, a spatial variation loss) to address the smoothness issue. These networks usually produce non-invertible transformations with ``folding'' in multiple voxel locations, indicated by a negative determinant of the Jacobian matrix of the transformation. While using a loss function that specifically penalizes the folding is a straightforward solution, this usually requires carefully tuning the regularization strength, especially when there are also other losses. In this paper we address this problem from a different angle, by investigating possible training mechanisms that will help the network avoid negative Jacobians and produce smoother deformations. We contribute two independent ideas in this direction. Both ideas greatly reduce the number of folding locations in the predicted deformation, without making changes to the hyperparameters or the architecture used in the existing baseline registration network.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper proposes two independent training mechanisms for unsupervised deep registration networks to reduce the number of locations with negative Jacobian determinants (folding) in the predicted deformation fields. The central claim is that these mechanisms achieve substantial reductions in folding while requiring no changes to the hyperparameters or architecture of a baseline registration network that uses standard similarity and smoothness losses.

Significance. If the mechanisms can be shown to deliver the claimed reductions on standard medical datasets while truly preserving baseline hyperparameters, the work would address a practical barrier to deploying deep registration models in settings that require diffeomorphic transformations. This could reduce reliance on post-hoc folding corrections or extensive loss-weight tuning.

major comments (1)
  1. [Abstract] Abstract: The assertion that the two ideas reduce folding 'without making changes to the hyperparameters or the architecture used in the existing baseline registration network' is load-bearing for the contribution. Any new training mechanism (auxiliary loss, sampling strategy, or regularizer) must be combined with the existing similarity + smoothness losses; the relative weighting coefficients are hyperparameters. The manuscript supplies no evidence that these coefficients remain fixed at their baseline values, remain optimal, or are unnecessary once the new mechanisms are present.
minor comments (1)
  1. [Abstract] Abstract: The two proposed ideas are not named or described, preventing assessment of whether they are genuinely independent of the baseline losses or architecture.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed review and the opportunity to clarify the central claim in the abstract. We address the major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The assertion that the two ideas reduce folding 'without making changes to the hyperparameters or the architecture used in the existing baseline registration network' is load-bearing for the contribution. Any new training mechanism (auxiliary loss, sampling strategy, or regularizer) must be combined with the existing similarity + smoothness losses; the relative weighting coefficients are hyperparameters. The manuscript supplies no evidence that these coefficients remain fixed at their baseline values, remain optimal, or are unnecessary once the new mechanisms are present.

    Authors: We confirm that the two proposed training mechanisms were added while keeping the similarity and smoothness loss weights exactly equal to the values used for the baseline network (i.e., no retuning or re-optimization of those coefficients was performed). The mechanisms are therefore compatible with the standard loss combination at its original hyperparameter settings. To address the lack of explicit evidence, the revised manuscript will include a dedicated paragraph and table that lists the precise loss weights employed for both the baseline and the augmented training runs, confirming they are identical. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical mechanisms proposed without self-referential derivation

full rationale

The paper proposes two training mechanisms to reduce negative Jacobian determinants in deep registration networks. The abstract states these ideas are independent, reduce folding locations, and require no changes to baseline hyperparameters or architecture. No equations, fitted parameters renamed as predictions, self-citations as load-bearing premises, or ansatzes smuggled via prior work are present in the provided text. The central claims rest on empirical outcomes on medical datasets rather than any derivation that reduces to its own inputs by construction. This is the expected non-finding for a methods paper focused on practical training adjustments.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no free parameters, axioms, or invented entities are identifiable from the provided text.

pith-pipeline@v0.9.0 · 5723 in / 1046 out tokens · 28928 ms · 2026-05-25T13:23:07.334555+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MorphSeek: Fine-grained Latent Representation-Level Policy Optimization for Deformable Image Registration

    cs.CV 2025-11 unverdicted novelty 7.0

    MorphSeek advances a latent representation-level policy optimization approach for deformable image registration that reports Dice gains on brain MRI, liver CT, and MR-CT benchmarks with high label efficiency.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · cited by 1 Pith paper · 5 internal anchors

  1. [1]

    Neuroimage 54(3), 2033–2044 (2011)

    Avants, B.B., Tustison, N.J., Song, G., Cook, P.A., Klein, A., Gee, J.C.: A repro- ducible evaluation of ants similarity metric performance in brain image registration. Neuroimage 54(3), 2033–2044 (2011)

    work page 2033

  2. [2]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J., Dalca, A.V.: An unsuper- vised learning model for deformable medical image registration. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 9252–9260 (2018) 8 D. Kuang et al

    work page 2018

  3. [3]

    In: Advances in neural information processing systems

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural information processing systems. pp. 2672–2680 (2014)

    work page 2014

  4. [4]

    In: Advances in neural information processing systems

    Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. In: Advances in neural information processing systems. pp. 2017–2025 (2015)

    work page 2017

  5. [5]

    Adam: A Method for Stochastic Optimization

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  6. [6]

    Frontiers in neuroscience 6, 171 (2012)

    Klein, A., Tourville, J.: 101 labeled brain images and a consistent human cortical labeling protocol. Frontiers in neuroscience 6, 171 (2012)

    work page 2012

  7. [7]

    Non-rigid image registration using fully convolutional networks with deep self-supervision

    Li, H., Fan, Y.: Non-rigid image registration using fully convolutional networks with deep self-supervision. arXiv preprint arXiv:1709.00799 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  8. [8]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Roh´ e, M.M., Datar, M., Heimann, T., Sermesant, M., Pennec, X.: Svf-net: Learning deformable image registration using shape matching. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 266–274. Springer (2017)

    work page 2017

  9. [9]

    In: International Conference on Medical image computing and computer-assisted intervention

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

    work page 2015

  10. [10]

    Unsupervised End-to-end Learning for Deformable Medical Image Registration

    Shan, S., Guo, X., Yan, W., Chang, E.I., Fan, Y., Xu, Y., et al.: Unsupervised end-to-end learning for deformable medical image registration. arXiv preprint arXiv:1711.08608 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  11. [11]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Sokooti, H., de Vos, B., Berendsen, F., Lelieveldt, B.P., Iˇ sgum, I., Staring, M.: Nonrigid image registration using multi-scale 3d convolutional neural networks. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 232–239. Springer (2017)

    work page 2017

  12. [12]

    In: Deep Learning for Medical Image Analysis, pp

    Wang, S., Kim, M., Wu, G., Shen, D.: Scalable high performance image registra- tion framework by unsupervised deep feature representations learning. In: Deep Learning for Medical Image Analysis, pp. 245–269. Elsevier (2017)

    work page 2017

  13. [13]

    Empirical Evaluation of Rectified Activations in Convolutional Network

    Xu, B., Wang, N., Chen, T., Li, M.: Empirical evaluation of rectified activations in convolutional network. arXiv preprint arXiv:1505.00853 (2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  14. [14]

    NeuroImage 158, 378–396 (2017)

    Yang, X., Kwitt, R., Styner, M., Niethammer, M.: Quicksilver: Fast predictive image registration–a deep learning approach. NeuroImage 158, 378–396 (2017)

    work page 2017

  15. [15]

    In: Deep Learning in Medical Image Analysis and Multi- modal Learning for Clinical Decision Support, pp

    Yoo, I., Hildebrand, D.G., Tobin, W.F., Lee, W.C.A., Jeong, W.K.: ssemnet: Serial- section electron microscopy image registration using a spatial transformer network with learned features. In: Deep Learning in Medical Image Analysis and Multi- modal Learning for Clinical Decision Support, pp. 249–257. Springer (2017)

    work page 2017

  16. [16]

    Inverse-Consistent Deep Networks for Unsupervised Deformable Image Registration

    Zhang, J.: Inverse-consistent deep networks for unsupervised deformable image registration. arXiv preprint arXiv:1809.03443 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  17. [17]

    arXiv preprint (2017)

    Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. arXiv preprint (2017)

    work page 2017