Pith. sign in

REVIEW 4 major objections 5 minor 26 references

An Interpretable Representation Learning Approach for Diffusion Tensor Imaging

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A 9×9 grid of tract-level fractional anisotropy values, embedded by a β-TCVAE, gives an interpretable DTI representation that beats the best 1D baseline by 12.64% F1 on sex classification.

desk verdict A plausible new 2D tract representation for DTI, undercut by an unspecified evaluation protocol and a sex/group confound; the idea deserves referee time, but the empirical claim is not yet demonstrated. read the letter →

arxiv 2505.19110 v2 pith:JUJBATS4 submitted 2025-05-25 cs.CV cs.AIcs.LG

Vishwa Mohan Singh , Alberto Gaston Villagran Asiares , Luisa Sophie Schuhmacher , Kate Rendall , Simon Wei{ss}brod , David Rügamer , Inga Körte This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords diffusiontensorimagingtractographyfractionalanisotropyrepresentationlearningvariationalautoencoderdisentanglementsexclassificationinterpretability
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

Diffusion tensor imaging (DTI) tractography is rich in structural detail but awkward for deep learning: flattened 1D feature vectors lose spatial context, while working directly on 3D fiber geometry is hard to interpret. This paper tries to establish a middle path: arrange the mean fractional anisotropy (FA) values of 74 white-matter tracts into a fixed 9×9 grayscale image, and learn a latent embedding from it with a β-TCVAE whose decoder is spatially broadcast. The paper reports that, for a downstream sex-classification task, this 2D representation raises F1 by 12.64 percentage points over the best 1D baseline (68.78 to 81.37), stays within reach of the 3D autoencoder (82.06 F1), and produces a more disentangled latent space as measured by the Mutual Information Gap (0.0535 vs 0.0344). Because the grid positions correspond to anatomy, the resulting latent dimensions and SHAP attributions can be read back as tract-level explanations rather than uninterpretable geometry.

What carries the argument

The load-bearing object is the fixed 9×9 fractional-anisotropy grid. Each of the 74 cells holds one tract's mean FA value; the cells are placed by multidimensional scaling so that nearby tracts in the brain stay nearby in the image, and the Hungarian algorithm resolves collisions so every subject uses the same grid coordinates. The grid is then fed through a β-total-correlation VAE with a spatial broadcast decoder: the β-weighted total-correlation term penalizes statistical dependence among latent dimensions (disentanglement), and the broadcast decoder forces the latent code to explain spatially structured patterns rather than memorizing pixel positions. This machinery converts raw tractography into a compact, comparable, anatomy-aligned input for standard convolutional and representation-learning models.

What would settle it

Re-run the same 2D β-TCVAE and classifier with cohort group (athlete/control) included as a covariate or with cross-validation stratified by group; if the F1 advantage over the 1D baseline largely disappears within homogeneous groups, the reported gain is driven by the sex/group confound rather than by the spatial grid representation.

Watch

Extended reading notes

Core claim

The central claim is that a 2D spatial layout of tract-level FA values is a better inductive bias for DTI representation learning than a 1D vector or raw 3D tractography point clouds. Concretely, the authors convert 74 tract-averaged FA values into a 9×9 grayscale image (positions fixed by multidimensional scaling plus Hungarian assignment), encode it with a β-TCVAE whose latent mean has 32 dimensions, and decode with a spatial broadcast decoder. With an auxiliary classifier on the latent mean, the 2D β-TCVAE achieves 81.37 F1 on sex classification compared with 68.78 for the best 1D grouped DNN, while the 3D VAE achieves 82.06; the 2D model's KNN separability is 83.33 and its MIG is 0.0535, above the 3D model's 0.0344. The paper therefore concludes that the 2D representation preserves enough spatial information to match 3D performance while being substantially more interpretable and more disentangled.

Load-bearing premise

The sex-classification result assumes that the classifier is learning biological sex rather than the athlete/control grouping, because the soccer group has 46 males and 23 females while the control group has 11 males and 25 females, and the model does not adjust for group membership.

Editorial extensions

If this is right

  • The 2D grid gives a practical middle ground: a DTI representation that supports standard 2D deep learning and anatomy-level interpretation while matching 3D accuracy within the reported variance.
  • A researcher can reuse the same fixed grid across subjects and modalities, enabling late-fusion architectures that combine DTI with other MRI sequences.
  • The disentangled latent space (MIG 0.0535) and high KNN separability (83.33) indicate the 32-dimensional representation retains class-relevant structure usable in small clinical cohorts.
  • SHAP attributions on the grid localize sex differences to specific tracts (left corona radiata, right superior longitudinal fasciculus, right corticospinal tract), giving a concrete anatomical readout rather than a black-box prediction.
  • If the grid is a stable preprocessing step, future models can treat tract-level FA images like ordinary images, opening the door to convolutions, attention, and pretraining on larger DTI datasets.

Reading between the lines

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

  • Beyond the paper: the same 9×9 layout could be computed for other tract-level diffusion measures (mean diffusivity, radial diffusivity, or fiber density) and stacked as channels, giving a multi-parameter DTI image without re-deriving the geometry.
  • Beyond the paper: a controlled ablation that scrambles the grid positions while keeping the same VAE would test whether the reported gains come from the anatomical layout or only from the spatial-broadcast architecture.
  • Beyond the paper: because the sex ratios differ between the athlete and control groups, a cohort-stratified or covariate-adjusted retraining could change the magnitude of the reported sex-classification gain.
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 / 5 minor

Summary. The paper proposes a 2D grid representation of DTI tractography, where 74 tract-level fractional anisotropy (FA) values are arranged into a 9x9 image via multidimensional scaling and the Hungarian algorithm. This image is fed into a beta-TCVAE with a spatial broadcast decoder to learn a disentangled latent embedding, which is then used for downstream sex classification and evaluated with auxiliary classification, triplet loss, and SimCLR objectives. The authors compare against 1D DNN and 3D autoencoder baselines and report that the 2D representation improves F1 over the best 1D model by 12.64% (15.74% in the abstract) and achieves higher mutual information gap (MIG) than the 3D VAE. The central claim is that the 9x9 FA grid encoded by beta-TCVAE yields a latent space that is both more separable and more disentangled than 1D or 3D alternatives.

Significance. If the evaluation were sound, the paper would make a useful contribution: compact and interpretable representations of tractography are valuable for multimodal and clinical analyses, and the use of a spatial broadcast decoder on a geometrically arranged tract grid is a sensible idea. The paper ships a public code repository, provides explicit pseudocode for the representation construction, and includes SHAP-based interpretation linked to known white-matter sexual dimorphism. However, the load-bearing empirical claims are not yet supported because the evaluation protocol is undocumented and the dataset contains a sex/group confound. The contribution is therefore promising but currently unverified.

major comments (4)
  1. [Section 3, Table 1] The manuscript does not state how models were trained and evaluated: there is no train/validation/test split, no cross-validation scheme, no repetition count, and no definition of the reported +/- intervals. Since the 2D beta-TCVAE models include an auxiliary classifier trained on only 105 subjects, the reported F1 values could reflect training-set memorization rather than generalization. The statement that the 2D representation 'improves over the best 1D model by 12.64% in F1 score' therefore requires a documented held-out evaluation before it can be assessed.
  2. [Section 2.1] Sex labels are confounded with athlete/control group membership: the soccer cohort has 46 male and 23 female subjects, while the control cohort has 11 male and 25 female subjects. A classifier trained to predict sex can instead separate cohorts, and the paper does not adjust for group or report within-group performance. This confound undermines the downstream classification results and the SHAP interpretation in Section 4. Please report stratified results, e.g., sex classification separately within each group, or a group-balanced evaluation.
  3. [Abstract vs. Section 4] The abstract states a 15.74% improvement, while Section 4 and Table 1 state 12.64%. These numbers are inconsistent, and the derivation of either from Table 1 is unclear: the largest 2D F1 value (81.42) minus the best 1D F1 (68.78) equals 12.64 percentage points, not 15.74%, and the relative improvement is 18.4%. The authors must reconcile this discrepancy and specify the exact comparison the headline number refers to.
  4. [Table 1] The differences between models are smaller than the reported standard deviations: 2D beta-TCVAE + Aux has F1 81.37 +/- 13.4, while 3D VAE + Aux has 82.06 +/- 9.7 and 1D Group DNN has 68.78 +/- 15.7. In the absence of significance testing, the claim of 'no significant drop from the 3D equivalent' is unsupported. Please provide repeated-run statistics with defined intervals and a significance test, or remove the claim.
minor comments (5)
  1. [Throughout] The notation 'FA' is typeset as 'F A' and 'VAE' as 'V AE' in several places; please fix the spacing for readability.
  2. [Section 2.1] The 9x9 grid has 81 cells for 74 tracts; please explain how empty cells are handled and whether the discretized grid is stable across subjects or computed per subject.
  3. [Section 3] The phrase 'the metrics from the best classifier from LazyClassifier' is undefined; specify which classifier was selected, its hyperparameters, and how selection bias was avoided.
  4. [Appendix E] Figure 2's axes and left/right panels are not explained in the text; add a caption that states what is plotted and how the polarity of interactions should be read.
  5. [Section 3, Table 1] The KNN separability metric is reported with k=3 but no description of how the KNN classifier was trained or evaluated; please clarify whether it was run on the same subjects used for training the autoencoder.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the reported F1/MIG gains are empirical comparisons against external baselines; the protocol concerns are validity issues, not circularity.

full rationale

The paper's central claim is an empirical comparison: a β-TCVAE with a spatial broadcast decoder is trained on a fixed 9x9 FA grid, and its downstream sex-classification F1 and MIG values are compared with 1D DNN and 3D AE baselines. The grid construction (MDS projection followed by Hungarian assignment) depends only on tract centroid geometry and does not use the sex labels or the reported metrics as input. The loss equations (Appendix C) are standard β-TCVAE, auxiliary-classifier, triplet, and SimCLR objectives; no fitted parameter is renamed as a prediction, and no quantity in the paper reduces by construction to another quantity it is claimed to predict. The paper invokes no uniqueness theorem from its own authors and contains no load-bearing self-citation. The serious concerns raised by the reader—absence of a described held-out split, class/group confounding, and the abstract/body discrepancy between 15.74% and 12.64%—are threats to evaluation validity and reproducibility, not circularity. They do not make the derivation equivalent to its inputs. Therefore the appropriate circularity finding is none.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The paper's central claims depend on multiple unreported hyperparameters and on assumptions about parcellation consistency, grid stability, label purity, and evaluation protocol. These are not fitted physical constants, but the F1 and MIG numbers cannot be independently replicated or interpreted without them.

free parameters (7)
  • Grid size (9x9) = 9x9
    Hand-chosen; 74 tracts are placed in 81 cells with empty cells allowed, and no comparison to other grid sizes is reported.
  • Latent dimension = 32
    Hand-chosen; no ablation on latent size is reported, and the central latent-space claims depend on it.
  • Beta (TC-VAE weight) = not reported
    Controls the total-correlation penalty in Eq. 3 and therefore influences the MIG/disentanglement results.
  • lambda_VAE, lambda_Cls, lambda_Triplet, lambda_SimCLR = not reported
    Loss weights in Eqs. 4, 5, and 7 are not reported; they directly affect every F1 and MIG value in Table 1.
  • Triplet margin alpha = not reported
    Appears in Eq. 6; not specified.
  • SimCLR temperature tau = not reported
    Appears in Eq. 8; not specified.
  • KNN k (separability metric) = 3
    Used for the Sep metric; no sensitivity analysis is provided.
assumptions (6)
  • domain assumption The WMA800 atlas parcellation labels the same 74 anatomical tracts consistently across all subjects.
    Section 2.1; if tract labels are inconsistent across subjects, the grid positions and FA values are not comparable.
  • ad hoc to paper MDS plus Hungarian assignment preserves enough spatial structure for the VAE and is stable across subjects.
    Algorithm 1; no baseline with a randomized grid or a stability analysis is reported.
  • domain assumption Tract-level mean FA captures microstructure relevant to sex classification.
    Section 2.1; FA values are averaged per tract, ignoring within-tract distributions and other diffusion measures.
  • domain assumption Sex labels are not confounded with athlete/control group membership.
    Section 2.1; 46M/23F soccer players vs 11M/25F controls, with no adjustment for group.
  • standard math The beta-TCVAE loss decomposition into mutual information, total correlation, and dimension-wise KL is a valid disentanglement objective.
    Appendix C Eq. 3, based on Chen et al. 2018; accepted as a starting point.
  • domain assumption The evaluation uses a proper held-out or cross-validated protocol with no leakage.
    Section 3 does not specify train/test splits, seeds, or whether MDS is fit on the whole dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Interpretable Representation Learning Approach for Diffusion Tensor Imaging." pith.science (2026). https://pith.science/paper/JUJBATS4

@misc{pith2026250519110,
  author       = {Pith},
  title        = {Pith review of: An Interpretable Representation Learning Approach for Diffusion Tensor Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUJBATS4}},
  note         = {Machine review of arXiv:2505.19110}
}
read the original abstract

Diffusion Tensor Imaging (DTI) tractography offers detailed insights into the structural connectivity of the brain, but presents challenges in effective representation and interpretation in deep learning models. In this work, we propose a novel 2D representation of DTI tractography that encodes tract-level fractional anisotropy (FA) values into a 9x9 grayscale image. This representation is processed through a Beta-Total Correlation Variational Autoencoder with a Spatial Broadcast Decoder to learn a disentangled and interpretable latent embedding. We evaluate the quality of this embedding using supervised and unsupervised representation learning strategies, including auxiliary classification, triplet loss, and SimCLR-based contrastive learning. Compared to the 1D Group deep neural network (DNN) baselines, our approach improves the F1 score in a downstream sex classification task by 15.74% and shows a better disentanglement than the 3D representation.

Figures

Figures reproduced from arXiv: 2505.19110 by the authors.

Figure 1
Figure 1. Architecture to convert the Dense DTI representation to an interpretable latent vector Z. The representation learning and classification are performed on the estimated mean mu. The reconstruction shows latent values, which are the ex￾plainers for each tract. 2.2. Autoencoder and Representation Learning The core of our model is a variational autoencoder (Kingma et al., 2013). The encoder compresses the 9×9 FA image i… view at source ↗
Figure 2
Figure 2. Graph that shows the impact of a tract (right) and the polarity of the interaction with the target i.e. sex classification (left). 10 [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 20 canonical work pages

  1. [1]

    Cerebral white matter myelination and relations to age, gender, and cognition: a selective review

    Irina S Buyanova and Marie Arsalidou. Cerebral white matter myelination and relations to age, gender, and cognition: a selective review. Frontiers in human neuroscience, 15: 0 662031, 2021

    work page 2021

  2. [2]

    Isolating sources of disentanglement in variational autoencoders

    Ricky TQ Chen, Xuechen Li, Roger B Grosse, and David K Duvenaud. Isolating sources of disentanglement in variational autoencoders. Advances in neural information processing systems, 31, 2018

    2018

  3. [3]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597--1607. PmLR, 2020

    2020

  4. [4]

    Tractgraphcnn: anatomically informed graph cnn for classification using diffusion mri tractography

    Yuqian Chen, Fan Zhang, Leo R Zekelman, Tengfei Xue, Chaoyi Zhang, Yang Song, Nikos Makris, Yogesh Rathi, Weidong Cai, and Lauren J O’Donnell. Tractgraphcnn: anatomically informed graph cnn for classification using diffusion mri tractography. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1--5. IEEE, 2023

    work page 2023

  5. [5]

    A separability-based approach to quantifying generalization: which layer is best?

    Luciano Dyballa, Evan Gerritz, and Steven W Zucker. A separability-based approach to quantifying generalization: which layer is best? arXiv preprint arXiv:2405.01524, 2024

    work page Pith review arXiv 2024

  6. [6]

    Variational autoencoders for generating synthetic tractography-based bundle templates in a low-data setting

    Yixue Feng, Bramsh Q Chandio, Sophia I Thomopoulos, Tamoghna Chattopadhyay, and Paul M Thompson. Variational autoencoders for generating synthetic tractography-based bundle templates in a low-data setting. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1--6. IEEE, 2023

    work page 2023

  7. [7]

    Deep metric learning using triplet network

    Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-based pattern recognition: third international workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84--92. Springer, 2015

    work page 2015

  8. [8]

    Andrei Irimia, Xiaoyu Lei, Carinna M Torgerson, Zachary J Jacokes, Sumiko Abe, and John D Van Horn. Support vector machines, multidimensional scaling and magnetic resonance imaging reveal structural brain abnormalities associated with the interaction between autism spectrum disorder and sex. Frontiers in computational neuroscience, 12: 0 93, 2018

    work page 2018

Show all 26 references
  1. [9]

    Design and validation of diffusion mri models of white matter

    Ileana O Jelescu and Matthew D Budde. Design and validation of diffusion mri models of white matter. Frontiers in physics, 5: 0 61, 2017

    2017

  2. [10]

    Auto-encoding variational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013

    2013

  3. [11]

    Trafic: fiber tract classification using deep learning

    Prince D Ngattai Lam, Gaetan Belhomme, Jessica Ferrall, Billie Patterson, Martin Styner, and Juan C Prieto. Trafic: fiber tract classification using deep learning. In Medical Imaging 2018: Image Processing, volume 10574, pages 257--265. SPIE, 2018

    2018

  4. [12]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017

    2017

  5. [13]

    Review of the development of multidimensional scaling methods

    Al Mead. Review of the development of multidimensional scaling methods. Journal of the Royal Statistical Society: Series D (The Statistician), 41 0 (1): 0 27--39, 1992

    1992

  6. [14]

    Men and women are different: diffusion tensor imaging reveals sexual dimorphism in the microstructure of the thalamus, corpus callosum and cingulum

    K Menzler, M Belke, E Wehrmann, K Krakow, U Lengler, Andreas Jansen, Hajo M Hamer, Wolfgang H Oertel, Felix Rosenow, and Susanne Knake. Men and women are different: diffusion tensor imaging reveals sexual dimorphism in the microstructure of the thalamus, corpus callosum and ci...

    2011

  7. [15]

    Automatic tractography segmentation using a high-dimensional white matter atlas

    Lauren J O'Donnell and Carl-Fredrik Westin. Automatic tractography segmentation using a high-dimensional white matter atlas. IEEE transactions on medical imaging, 26 0 (11): 0 1562--1575, 2007

    2007

  8. [16]

    Unbiased groupwise registration of white matter tractography

    Lauren J O’Donnell, William M Wells III, Alexandra J Golby, and Carl-Fredrik Westin. Unbiased groupwise registration of white matter tractography. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 123--130. Springer, 2012

    2012

  9. [17]

    Pandala and Bruno Silva

    S.R. Pandala and Bruno Silva. Lazy predict project. https://pypi.org/project/lazypredict/, May 2019. Python package for quick evaluation of multiple machine learning models

    2019

  10. [18]

    Integrated brain connectivity analysis with fmri, dti, and smri powered by interpretable graph neural networks

    Gang Qu, Ziyu Zhou, Vince D Calhoun, Aiying Zhang, and Yu-Ping Wang. Integrated brain connectivity analysis with fmri, dti, and smri powered by interpretable graph neural networks. ArXiv, pages arXiv--2408, 2024

    2024

  11. [19]

    Factorization machines

    Steffen Rendle. Factorization machines. In 2010 IEEE International conference on data mining, pages 995--1000. IEEE, 2010

    2010

  12. [20]

    Diffusion tensor imaging of the brain: review of clinical applications

    Pia C Sundgren, Q Dong, D Gomez-Hassan, SK Mukherji, P Maly, and R Welsh. Diffusion tensor imaging of the brain: review of clinical applications. Neuroradiology, 46: 0 339--350, 2004

    2004

  13. [21]

    The role of spatial embedding in mouse brain networks constructed from diffusion tractography and tracer injections

    Scott Trinkle, Sean Foxley, Gregg Wildenberg, Narayanan Kasthuri, and Patrick La Rivi \`e re. The role of spatial embedding in mouse brain networks constructed from diffusion tractography and tracer injections. Neuroimage, 244: 0 118576, 2021

    2021

  14. [22]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  15. [23]

    Spatial broadcast decoder: A simple architecture for learning disentangled representations in vaes

    Nicholas Watters, Loic Matthey, Christopher P Burgess, and Alexander Lerchner. Spatial broadcast decoder: A simple architecture for learning disentangled representations in vaes. arXiv preprint arXiv:1901.07017, 2019

    1901 arXiv

  16. [24]

    An anatomically curated fiber clustering white matter atlas for consistent white matter tract parcellation across the lifespan

    Fan Zhang, Ye Wu, Isaiah Norton, Laura Rigolo, Yogesh Rathi, Nikos Makris, and Lauren J O'Donnell. An anatomically curated fiber clustering white matter atlas for consistent white matter tract parcellation across the lifespan. Neuroimage, 179: 0 429--447, 2018

    2018

  17. [25]

    Deep white matter analysis (deepwma): Fast and consistent tractography segmentation

    Fan Zhang, Suheyla Cetin Karayumak, Nico Hoffmann, Yogesh Rathi, Alexandra J Golby, and Lauren J O’Donnell. Deep white matter analysis (deepwma): Fast and consistent tractography segmentation. Medical image analysis, 65: 0 101761, 2020

    2020

  18. [26]

    Disentangled and proportional representation learning for multi-view brain connectomes

    Yanfu Zhang, Liang Zhan, Shandong Wu, Paul Thompson, and Heng Huang. Disentangled and proportional representation learning for multi-view brain connectomes. In Medical Image Computing and Computer Assisted Intervention--MICCAI 2021: 24th International Conference, Strasbourg, F...

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.