pith. sign in

arxiv: 2604.07154 · v1 · submitted 2026-04-08 · 💻 cs.CV · cs.AI

Bridging MRI and PET physiology: Untangling complementarity through orthogonal representations

Sonja Adomeit , Kartikay Tehlan , Lukas F\"orner , Katharina Weisser , Helen Scholtiseek , David Kaufmann , Julie Steinestel , Constantin Lapa
show 2 more authors
Thomas Kr\"oncke Thomas Wendler
This is my paper

Pith reviewed 2026-05-10 17:43 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords PSMA PETmultiparametric MRIorthogonal subspace decompositionimplicit neural representationsprostate cancermultimodal imagingmodality complementaritysubspace separation
0
0 comments X

The pith

PSMA PET contains signal components not recoverable from MRI-derived physiological descriptors, shown by the largest orthogonal residual in tumor regions.

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

This paper develops a subspace decomposition to separate PSMA PET uptake into an MRI-explainable physiological envelope and an orthogonal residual that lies outside the span of MRI features. It trains an intensity-based implicit neural representation to map MRI vectors to PET values while applying singular value decomposition projection to penalize any non-orthogonal components, enforcing separation between tissue properties such as structure, diffusion, and perfusion and intracellular PSMA expression. Tested on multiparametric MRI and PET scans from 13 prostate cancer patients, the method absorbs shared signal into the learned envelope and leaves the largest residual in tumor areas. This shows that PET supplies information on PSMA uptake that cannot be recovered from MRI alone, clarifying when the two modalities provide genuinely complementary data for imaging decisions.

Core claim

The paper establishes that PSMA PET uptake decomposes into an MRI-explainable physiological envelope, learned by a non-spatial implicit neural representation from multiparametric MRI feature vectors, plus an orthogonal residual isolated by singular value decomposition projection regularization. This residual reflects signal components not expressible within the MRI manifold and is shown to be largest in tumor regions across 13 patients, indicating that PET captures irreducible aspects of intracellular PSMA expression beyond MRI-measurable physiology.

What carries the argument

Subspace decomposition framework that uses singular value decomposition projection regularization on an implicit neural representation to enforce orthogonality between the MRI feature manifold and the PET residual.

If this is right

  • Residuals spanned by MRI features are fully absorbed into the learned physiological envelope.
  • The orthogonal residual being largest in tumors demonstrates that PSMA PET supplies unique information beyond MRI descriptors.
  • The decomposition supplies a geometric characterization of modality complementarity for prostate cancer imaging.
  • Rational acquisition strategies can prioritize PET where the orthogonal component is expected to be large.

Where Pith is reading between the lines

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

  • The size of the orthogonal residual could be used as a per-patient metric to assess whether PET adds diagnostic value beyond MRI.
  • The same orthogonal-separation logic might be applied to other modality pairs to quantify their unique contributions without image translation.
  • Correlation of the residual map with molecular PSMA expression assays would test whether the independent component tracks the intended biology.

Load-bearing premise

The MRI feature manifold fully spans every physiological property that could account for PET uptake, so that the SVD projection cleanly isolates independent components without discarding clinically relevant shared signal.

What would settle it

A statistical test showing that the magnitude of the orthogonal residual does not differ significantly between tumor and non-tumor regions or that the residual remains correlated with MRI features after projection.

Figures

Figures reproduced from arXiv: 2604.07154 by Constantin Lapa, David Kaufmann, Helen Scholtiseek, Julie Steinestel, Kartikay Tehlan, Katharina Weisser, Lukas F\"orner, Sonja Adomeit, Thomas Kr\"oncke, Thomas Wendler.

Figure 1
Figure 1. Figure 1: Graphical abstract. Seven MRI sequences are used to synthesize a PSMA [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of a patient case featuring: (a) the squared parallel residual; [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of orthogonal and parallel error metrics, stratified by tissue [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
read the original abstract

Multimodal imaging analysis often relies on joint latent representations, yet these approaches rarely define what information is shared versus modality-specific. Clarifying this distinction is clinically relevant, as it delineates the irreducible contribution of each modality and informs rational acquisition strategies. We propose a subspace decomposition framework that reframes multimodal fusion as a problem of orthogonal subspace separation rather than translation. We decompose Prostate-Specific Membrane Antigen (PSMA) PET uptake into an MRI-explainable physiological envelope and an orthogonal residual reflecting signal components not expressible within the MRI feature manifold. Using multiparametric MRI, we train an intensity-based, non-spatial implicit neural representation (INR) to map MRI feature vectors to PET uptake. We introduce a projection-based regularization using singular value decomposition to penalize residual components lying within the span of the MRI feature manifold. This enforces mathematical orthogonality between tissue-level physiological properties (structure, diffusion, perfusion) and intracellular PSMA expression. Tested on 13 prostate cancer patients, the model demonstrates that residual components spanned by MRI features are absorbed into the learned envelope, while the orthogonal residual is largest in tumour regions. This indicates that PSMA PET contains signal components not recoverable from MRI-derived physiological descriptors. The resulting decomposition provides a structured characterization of modality complementarity grounded in representation geometry rather than image translation.

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

3 major / 2 minor

Summary. The paper proposes a subspace decomposition framework that uses an intensity-based non-spatial implicit neural representation (INR) to map multiparametric MRI features to PSMA PET uptake, combined with SVD-based projection regularization to enforce orthogonality. This separates PET signal into an MRI-explainable physiological envelope and an orthogonal residual, which the authors interpret as reflecting irreducible components such as intracellular PSMA expression. Experiments on 13 prostate cancer patients show the orthogonal residual is largest in tumor regions, supporting the claim that PSMA PET contains signal components not recoverable from MRI-derived physiological descriptors.

Significance. If validated, the geometric approach to quantifying shared versus modality-specific information could inform rational multimodal imaging strategies in prostate cancer by clarifying irreducible contributions of each modality. The INR-plus-SVD construction provides a structured alternative to standard latent fusion methods, with potential for broader application in disentangling physiological signals across imaging modalities.

major comments (3)
  1. [Methods] The SVD projection regularization (described in the Methods) enforces orthogonality by construction between the learned MRI feature manifold and the residual; therefore the observation that the residual is largest in tumors (Abstract) partly follows from the regularization choice rather than from independent measurement of biological complementarity.
  2. [Methods] The INR is explicitly non-spatial and intensity-based; the manuscript reports no ablation that adds spatial coordinates or compares against a spatially aware baseline, so the tumor enrichment of the orthogonal residual could be explained by omitted spatial heterogeneity (e.g., boundary effects or local context visible in the same mpMRI volumes) rather than true physiological independence.
  3. [Results] The central demonstration on 13 patients states that the residual is largest in tumor regions, yet the abstract and results summary provide no quantitative metrics, error bars, cross-validation details, or statistical tests; this weakens the strength of the claim that the residual reveals irreducible biology.
minor comments (2)
  1. [Abstract] The abstract would be strengthened by including at least one key quantitative result (e.g., mean residual difference or correlation coefficient) and a brief statement of statistical evaluation.
  2. [Methods] Notation for the MRI feature manifold and the projection operator should be introduced with explicit equations early in the Methods to improve readability for readers unfamiliar with INR-based mapping.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments. We address each major point below, clarifying our methodological choices while acknowledging limitations and outlining targeted revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Methods] The SVD projection regularization (described in the Methods) enforces orthogonality by construction between the learned MRI feature manifold and the residual; therefore the observation that the residual is largest in tumors (Abstract) partly follows from the regularization choice rather than from independent measurement of biological complementarity.

    Authors: We agree that the SVD projection enforces orthogonality by mathematical construction. However, the mapping itself is learned from data via the INR, so the residual represents the component of observed PET uptake that cannot be expressed as a function of the MRI feature vectors. The spatial variation of this residual—being larger in tumors—is an empirical outcome of the learned decomposition rather than a predetermined result of the regularizer. We will revise the Methods and Discussion sections to explicitly distinguish the enforced orthogonality from the data-driven observation of its spatial distribution, thereby clarifying that the tumor enrichment reflects biological complementarity rather than circularity. revision: partial

  2. Referee: [Methods] The INR is explicitly non-spatial and intensity-based; the manuscript reports no ablation that adds spatial coordinates or compares against a spatially aware baseline, so the tumor enrichment of the orthogonal residual could be explained by omitted spatial heterogeneity (e.g., boundary effects or local context visible in the same mpMRI volumes) rather than true physiological independence.

    Authors: The non-spatial, intensity-only INR was intentionally selected to isolate complementarity at the level of tissue physiological descriptors (e.g., diffusion, perfusion) without spatial context, thereby testing whether PET signal contains components irreducible to these descriptors. We acknowledge that unmodeled spatial heterogeneity could contribute to the observed residual pattern and that no spatial ablation was performed. In revision we will expand the Methods to justify this design choice and add a dedicated limitations paragraph discussing potential spatial confounds, along with a sensitivity analysis if data permits. revision: partial

  3. Referee: [Results] The central demonstration on 13 patients states that the residual is largest in tumor regions, yet the abstract and results summary provide no quantitative metrics, error bars, cross-validation details, or statistical tests; this weakens the strength of the claim that the residual reveals irreducible biology.

    Authors: We concur that the current Results lack the quantitative rigor needed to support the central claim. The revised manuscript will report mean residual values (with standard deviations) for tumor versus non-tumor regions, leave-one-patient-out cross-validation performance, and statistical comparisons (paired t-tests or Wilcoxon signed-rank tests with p-values and effect sizes). These metrics will be added to the Results section, with a concise summary incorporated into the Abstract. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's derivation decomposes PET uptake via an intensity-based INR fitted to MRI features, with SVD projection regularization enforcing that the residual lies outside the MRI feature span. The central empirical result—that this orthogonal residual is largest in tumor regions—is measured directly on the 13-patient cohort after training and is not equivalent to the method inputs by construction. The claim of modality complementarity follows from the observed spatial pattern in the data rather than reducing tautologically to the orthogonality constraint or any fitted parameter. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing steps in the provided text. The approach remains self-contained against the external patient data benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the assumption that MRI features form a complete manifold for physiological explanation and that SVD projection isolates independent biological signals; no explicit free parameters or new entities are named in the abstract, but the orthogonality enforcement is definitional.

axioms (1)
  • domain assumption MRI feature vectors span all physiological properties relevant to PET uptake
    Invoked when training the INR to map MRI to PET and interpreting residuals as non-recoverable.

pith-pipeline@v0.9.0 · 5565 in / 1314 out tokens · 77879 ms · 2026-05-10T17:43:24.129471+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · 1 internal anchor

  1. [1]

    Advances in neural information processing systems29(2016)

    Bousmalis, K., Trigeorgis, G., Silberman, N., Krishnan, D., Erhan, D.: Domain separation networks. Advances in neural information processing systems29(2016)

    work page 2016

  2. [2]

    European Radiology pp

    Chen, Z., Bi, S., Shan, Y., Wang, F., Wang, Y., Qi, Z., Wang, T., Li, X., Li, S., Xiao, H., et al.: Mri-to-pet synthesis via deep learning for amyloid-βquantification in alzheimer’s disease. European Radiology pp. 1–13 (2026)

    work page 2026

  3. [3]

    European Urology Oncol- ogy (2025)

    Chow, K.M., Lee, A., Peh, D., Tan, Y.G., Tay, K.J., Ho, H., Cheng, C., Lam, W., Thang, S.P., Tuan, J., et al.: Combined prostate-specific membrane antigen positron emission tomography and multiparametric magnetic resonance imaging for the diagnosis of clinically significant prostate cancer. European Urology Oncol- ogy (2025)

    work page 2025

  4. [4]

    In: International conference on medical image computing and computer-assisted intervention

    Havaei, M., Guizard, N., Chapados, N., Bengio, Y.: Hemis: Hetero-modal im- age segmentation. In: International conference on medical image computing and computer-assisted intervention. pp. 469–477. Springer (2016)

    work page 2016

  5. [5]

    In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition

    Lee, M., Pavlovic, V.: Private-shared disentangled multimodal vae for learning of latent representations. In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition. pp. 1692–1700 (2021)

    work page 2021

  6. [6]

    The British journal of radiology95(1131), 20210728 (2022)

    Liu, F.Y., Sheng, T.W., Tseng, J.R., Yu, K.J., Tsui, K.H., Pang, S.T., Wang, L.J., Lin, G.: Prostate-specific membrane antigen (psma) fusion imaging in prostate cancer: Pet–ct vs pet–mri. The British journal of radiology95(1131), 20210728 (2022)

    work page 2022

  7. [7]

    Nature Reviews Urology13(4), 226–235 (Apr 2016).https://doi.org/10.1038/nrurol.2016.26,https://www.nature.com/ articles/nrurol.2016.26

    Maurer, T., Eiber, M., Schwaiger, M., Gschwend, J.E.: Current use of PSMA–PET in prostate cancer management. Nature Reviews Urology13(4), 226–235 (Apr 2016).https://doi.org/10.1038/nrurol.2016.26,https://www.nature.com/ articles/nrurol.2016.26

    work page doi:10.1038/nrurol.2016.26 2016

  8. [8]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Molaei, A., Aminimehr, A., Tavakoli, A., Kazerouni, A., Azad, B., Azad, R., Mer- hof, D.: Implicit neural representation in medical imaging: A comparative survey. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2381–2391 (2023)

    work page 2023

  9. [9]

    In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI)

    Moussaoui, Y., Mateus, D., Taheri, N., Moussaoui, S., Carlier, T., Stute, S.: Im- plicit neural representations for end-to-end pet reconstruction. In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI). pp. 1–5. IEEE (2025)

    work page 2025

  10. [10]

    Dickerson

    Rokuss, M., Kovacs, B., Kirchhoff, Y., Xiao, S., Ulrich, C., Maier-Hein, K.H., Isensee, F.: From FDG to PSMA: A hitchhiker’s guide to multitracer, multicen- ter lesion segmentation in PET/CT imaging.https://doi.org/10.48550/arXiv. 2409.09478,http://arxiv.org/abs/2409.09478

    work page internal anchor Pith review doi:10.48550/arxiv

  11. [11]

    In: Proceedings of the 34th International Conference on Neural Information Processing Systems

    Sitzmann, V., Martel, J.N.P., Bergman, A.W., Lindell, D.B., Wetzstein, G.: Im- plicit neural representations with periodic activation functions. In: Proceedings of the 34th International Conference on Neural Information Processing Systems. pp. 7462–7473. NIPS ’20, Curran Associates Inc., Red Hook, NY, USA (Dec 2020), https://dl.acm.org/doi/10.5555/3495724.3496350

    work page doi:10.5555/3495724.3496350 2020

  12. [12]

    In: Advances in Neural Information Processing Systems

    Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Sing- hal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7537–7547. Curran Associates, Inc. (2020)

    work page 2020

  13. [13]

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

    Tehlan, K., Wendler, T.: Physiological neural representation for personalised tracer kinetic parameter estimation from dynamic pet. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 491–500. Springer (2025) Bridging MRI and PET 11

    work page 2025

  14. [14]

    medRxiv (2025)

    Theodorou, B., Dadu, A., Avants, B., Nalls, M., Sun, J., Faghri, F.: Mri2pet: realistic pet image synthesis from mri for automated inference of brain atrophy and alzheimer’s. medRxiv (2025)

    work page 2025

  15. [15]

    Frontiers in Oncology12, 831429 (2022)

    Tsechelidis, I., Vrachimis, A.: Psma pet in imaging prostate cancer. Frontiers in Oncology12, 831429 (2022)

    work page 2022

  16. [16]

    TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images

    Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., Bach, M., Segeroth, M.: TotalSegmenta- tor: Robust segmentation of 104 anatomic structures in CT images. Radiology: Ar- tificial Intelligence5(5), e230024 (2023).https://doi.org/10.1148/ryai.230024, http://pubs.rsna.org/doi/10.1148/...

    work page doi:10.1148/ryai.230024 2023

  17. [17]

    Wei, S., Luo, Y., Wang, Y., Luo, C.: Robust multimodal learning via representation decoupling.In:Europeanconferenceoncomputervision.pp.38–54.Springer(2024)

    work page 2024

  18. [18]

    In: International Workshop on Simu- lation and Synthesis in Medical Imaging

    Yaakub, S.N., McGinnity, C.J., Clough, J.R., Kerfoot, E., Girard, N., Guedj, E., Hammers, A.: Pseudo-normal pet synthesis with generative adversarial networks for localising hypometabolism in epilepsies. In: International Workshop on Simu- lation and Synthesis in Medical Imaging. pp. 42–51. Springer (2019)

    work page 2019