Pith. sign in

REVIEW 4 major objections 4 minor 25 references

TROI: Cross-Subject Pretraining with Sparse Voxel Selection for Enhanced fMRI Visual Decoding

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read TROI's learned voxel masks beat manually annotated ROIs in fMRI decoding

desk verdict TROI's cross-subject pretraining with a trainable sparse voxel mask is a plausible and honest small-sample fMRI decoding contribution, but the headline superiority claim over MindEye2 is undercut by a possibly contaminated evaluation split and by reconstruction metrics that get worse across all subjects. read the letter →

arxiv 2502.00412 v2 pith:CLR5OVL4 submitted 2025-02-01 cs.CV

classification cs.CV
keywords fMRIvisualdecodingcross-subjectpretrainingvoxelselectionsparsemasktraininglearningraterewindingimageretrievalreconstruction
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

This paper tries to show that fMRI visual decoding for a new subject can work without expert-drawn regions of interest (ROIs) and with very little data. It proposes TROI, a two-stage pipeline that first pretrains a decoding backbone on several other subjects using their annotated ROIs, then learns a sparse 0-1 mask over a coarsely segmented brain region to select roughly 3,000 informative voxels, and finally fine-tunes only the input layer on the new subject's first hour of data. The reported experiments on the Natural Scenes Dataset compare TROI against the MindEye2 baseline running on annotated ROIs and find that the learned mask improves brain-retrieval accuracy and high-level reconstruction metrics while using far fewer voxels. A sympathetic reader would care because, if this is right, it removes a major practical bottleneck: each new subject would no longer need manual anatomical labeling, only a coarse brain scan and an hour of data.

What carries the argument

The central mechanism is the trainable region-of-interest mask: a 0-1 vector $M'$ applied to flattened fMRI voxels by a Hadamard product, learned in Stage 1 by optimizing the decoding loss plus an L1 penalty on a continuous weight mask $M$, then thresholding, Gaussian low-pass filtering, and binarizing under a voxel budget $V$. The mask does the work of the manual ROI by carving a coarse occipital segmentation down to about 3,000 voxels; the low-pass filter injects spatial locality as a prior, and Stage 2's learning-rate rewinding retrains the input layer on that fixed mask so the sparse structure, not the fine-tuned weights of a pretrained model, governs the input representation.

What would settle it

If a random 3,000-voxel subset of the same coarse brain region, under identical pretraining and Stage-2 rewinding, matches TROI's reported brain-retrieval accuracy on a held-out subject, then the learned voxel mask is not what carries the improvement over annotated ROIs.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that a data-driven voxel mask learned with sparse-mask training plus low-pass filtering can replace a manually annotated ROI for decoding visual stimuli from fMRI, and that cross-subject pretraining makes this feasible when the new subject contributes only one hour of data. The system first trains a shared residual MLP backbone and a diffusion-prior reconstruction head on multiple subjects using annotated ROIs; for the held-out subject it then trains a weighted mask under an L1 penalty, zeroes out small weights, smooths the mask with a Gaussian low-pass filter, and binarizes it to a fixed voxel budget of 3,000. Stage 2 retrains the subject-specific input layer from scratch at a rewound learning rate on the fixed mask. On subjects 1, 2, 5, and 7, the paper reports that this procedure beats annotated-ROI MindEye2 on brain retrieval (for example, subject 1: 85.3% versus 77.6%) and on two-way identification in reconstruction (85.9% versus 83.6%), while pixel-level similarity metrics stay mixed.

Load-bearing premise

The load-bearing premise is that the voxel mask learned from one hour of a new subject's data, after thresholding and smoothing, keeps every voxel that matters for decoding, because Stage 2 only refines the input layer and cannot recover information from voxels the mask discarded.

Editorial extensions

If this is right

  • For a new subject with no ROI labels and only one hour of fMRI data, decoding pipelines can start from a coarse brain segmentation and learn the voxel subset automatically, instead of waiting for expert annotation.
  • The subject-specific input layer is the only part retrained for a new subject, so per-subject adaptation cost is low and overfitting risk is reduced in small-sample settings.
  • With a voxel budget around 3,000, the method is reported to reach or exceed annotated-ROI performance on brain retrieval and high-level reconstruction, implying that annotated ROIs contain a large fraction of task-irrelevant voxels.
  • The mask can be recomputed for each subject, so individual differences in brain organization are handled by the data rather than by a fixed atlas.
  • Different voxel budgets are available (1,000 to 3,000), giving a practical knob between input dimensionality and decoding accuracy.

Reading between the lines

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

  • The paper's strongest reported result is on brain retrieval; a natural next test is whether the same mask-learning recipe transfers to other fMRI tasks or other coarse segmentations, since the method makes no visual-cortex-specific assumption beyond the initial coarse crop.
  • One implication the authors do not develop is that the learned mask could serve as a fast functional atlas for new subjects, potentially reducing the need for anatomical alignment when building brain-computer interfaces or clinical fMRI tools.
  • A testable extension would be to run the Stage 1 mask learner on data from a different scanner or resolution and check whether a fixed voxel budget still yields comparable decoding, which would probe the method's dependence on preprocessing and spatial smoothness assumptions.
  • The mixed pixel-level results leave open whether a mask tuned for retrieval and high-level semantics necessarily preserves the fine-grained spatial detail that pixel-correlation measures reward.
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 / 4 minor

Summary. The paper proposes TROI, a two-stage data-driven ROI selection method for fMRI visual decoding. In Stage 1, a sparse 0-1 voxel mask is learned on a target subject's small sample (the first hour of NSD data) by optimizing a task loss with LASSO regularization, thresholding, and Gaussian low-pass filtering under a voxel budget. In Stage 2, the TROI module is retrained from scratch on the fixed mask using a learning-rate rewinding strategy. The backbone and MLP projector are pretrained on cross-subject data (subjects 2, 5, and 7 in the main experiments, with one passage also listing subject 3). Evaluation is reported for subjects 1, 2, 5, and 7 on image/brain retrieval and reconstruction metrics, comparing against MindEye2 with annotated ROI masks. The paper claims that TROI surpasses MindEye2 on the same small-sample dataset.

Significance. If the claims hold, TROI would be a practically useful contribution: it removes the need for manually annotated ROIs in fMRI decoding, reduces the input voxel count dramatically (to 3,000 from tens of thousands), and demonstrates that cross-subject pretraining plus sparse mask learning can transfer to a new subject with only one hour of data. The paper is honest in reporting all metrics in Table 1, and the ablations in Table 2 isolate the contributions of pretraining, low-pass filtering, and learning-rate rewinding. These are genuine strengths. However, the headline claim of surpassing MindEye2 is not supported by the current evaluation protocol: the paper does not establish that the evaluation images are disjoint from the fine-tuning images, and the metric-level results are mixed, with TROI consistently worse on PixCorr and SSIM. The four-subject comparison also lacks any error bars or significance testing, so several reported gains are within plausible noise. The core idea is defensible, but the central quantitative claim needs substantially stronger experimental support.

major comments (4)
  1. [Section 3.1-3.2] The evaluation split is the load-bearing issue. Section 3.1 states that the model is fine-tuned using the first 1-hour data of the target subject and then 'finally evaluating on the full dataset of the target subject.' Section 3.2 states that in NSD every image is viewed three times during data collection. If the full-dataset evaluation includes the later repetitions of images that appeared in the first-hour fine-tuning set, those exact images are present in the retrieval candidate pool and as reconstruction targets, which can inflate retrieval and reconstruction metrics through memorization. The paper does not report whether the evaluation images are disjoint from the fine-tuning images by image ID, how the standard MindEye2 held-out protocol was adapted, or the degree of overlap. This must be fixed: the comparison should be repeated on an evaluation set that is disjoint from the fine-tuning set at the image level, and the overlap statistics should be reported.
  2. [Abstract and Section 3.3.2, Table 1] The abstract's claim that 'our voxel selection method surpasses the state-of-the-art method MindEye2' overstates Table 1. TROI is worse than the annotated-ROI baseline on PixCorr and SSIM for all four subjects (e.g., subject 1 PixCorr 0.206 vs. 0.235; SSIM 0.382 vs. 0.428) and worse on image retrieval for subject 1 (93.6% vs. 94.0%). The improvements are concentrated in brain retrieval and Incep. The textual claims in Section 3.3.2 and the Conclusion should be qualified to reflect the mixed metric-level results rather than claiming overall superiority.
  3. [Table 1] There are no error bars, confidence intervals, or significance tests, and only four subjects are evaluated. Several reported differences are very small, for example subject 5 image retrieval 67.3% vs. 66.9% and brain retrieval 47.7% vs. 47.0%. Without repeated runs or a statistical test, these differences are not distinguishable from noise. The paper should at least report variability across multiple seeds or a paired significance test over subjects, and should temper conclusions for metrics whose differences are within a plausible noise range.
  4. [Section 2.3 and Algorithm 1] Stage 2 fixes the binary mask M′ produced by Stage 1 and retrains the TROI module from scratch, but the manuscript is ambiguous about whether this retraining updates only the input layer or the whole backbone. If only the input layer is trained, any informative voxel discarded by the Stage-1 mask is unrecoverable, and the method's success depends entirely on the quality of a mask learned from roughly 15 epochs on one hour of data. The paper should clarify the scope of Stage-2 retraining and provide evidence on mask stability, for example by comparing masks learned with different training durations or showing that discarded voxels do not contain unique information needed for reconstruction.
minor comments (4)
  1. [Section 2.1 vs. Section 3.1] The pretraining subject list is inconsistent: Section 2.1 says the cross-subject backbone is trained using data from Subjects 2, 3, 5, and 7, while Section 3.1 says Subjects 2, 5, and 7. Please correct the discrepancy.
  2. [Abstract] The line 'Keywords First keyword · Second keyword · More' is a placeholder and should be replaced with actual keywords.
  3. [Section 3.3.2] The sentence 'we slightly outperform MindEye2 in small sample scenarios' is contradicted by the PixCorr and SSIM columns of Table 1; the discussion should acknowledge these metrics explicitly.
  4. [Section 2.2 and Algorithm 1] The hyperparameters of the Gaussian low-pass filter (kernel size and standard deviation), the threshold th, and the voxel budget V are not specified in the experiments section; please report them for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical system comparison against an external baseline, and its equations are constructive definitions rather than restatements of their own outputs.

full rationale

This is an empirical systems paper rather than a derivation chain. The central claim is a quantitative comparison against MindEye2, an external baseline with its own published results and code, and the ablations (pretraining, low-pass filtering, learning-rate rewinding, voxel budget) are independently configured experiments rather than restatements of fitted values. The learned voxel mask is a trained artifact, not a quantity that is defined in terms of the reported evaluation metric; Eq. (5)-(7) merely specify a sparsity-constrained optimization and a thresholding binarization, which do not reduce to the final retrieval or reconstruction scores by construction. Stage 2 retrains from the Stage-1 mask, but that is a standard two-stage fitting procedure, not a circular prediction. The paper does cite a prior rewinding study [21] to motivate Stage 2, but that is genuine external work by other authors, not a load-bearing self-citation. One non-circular caveat exists: Section 3.1/3.2 state that fine-tuning uses the first 1-hour data while evaluation uses the full NSD dataset, and NSD repeats each image three times; this raises a possible evaluation-contamination risk, but that is a benchmark-design concern, not definitional circularity. No equation in the paper is equivalent to its own input, and no reported result is a renamed fit.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method rests on several hand-chosen hyperparameters, most of which are unreported, and on four domain assumptions about occipital coverage, spatial smoothness, cross-subject transfer, and the applicability of pruning-style rewinding. No new physical entities are introduced.

free parameters (6)
  • voxel budget V = 3000 (also 1000 and 2000 in Table 3)
    Chosen by hand; determines mask size and drives the results in Table 3.
  • LASSO coefficient psi = not reported
    Controls the sparsity penalty in Eq. (6); its value is not given in the paper.
  • Binarization threshold th = not reported
    Used in Eq. (7) and Algorithm 1 to decide which voxels survive after filtering.
  • Gaussian low-pass filter size and standard deviation = not reported
    Defines the spatial smoothing applied to the learned mask; no kernel parameters are stated.
  • Loss balance epsilon = not reported
    Balances the CLIP loss and the diffusion prior loss in Eq. (4); no value is given.
  • Training hyperparameters (optimizer, learning rate, temperature tau) = not reported
    Central to both stages but absent from the experimental section.
assumptions (4)
  • domain assumption The coarse occipital-lobe segmentation contains all voxels needed for visual decoding.
    Input is restricted to a 60 by 40 by 40 grid primarily covering the occipital lobe (Section 3.2); informative voxels outside this crop cannot be selected.
  • domain assumption Functional brain regions are spatially local, so Gaussian smoothing of the learned mask is safe.
    Section 2.2 says the low-pass filter 'leverages the spatial locality of brain functional regions'; isolated but informative voxels could be removed by smoothing.
  • domain assumption A shared residual backbone trained on other subjects transfers to a new subject after a subject-specific linear input layer.
    The backbone is pretrained on Subjects 2, 5, and 7 and then applied to Subject 1 as a new subject (Section 3.1).
  • ad hoc to paper Learning rate rewinding from model pruning transfers to retraining an input-layer mask for fMRI data.
    Borrowed from pruning literature [21] without a derivation or independent validation on fMRI data beyond the reported experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TROI: Cross-Subject Pretraining with Sparse Voxel Selection for Enhanced fMRI Visual Decoding." pith.science (2026). https://pith.science/paper/CLR5OVL4

@misc{pith2026250200412,
  author       = {Pith},
  title        = {Pith review of: TROI: Cross-Subject Pretraining with Sparse Voxel Selection for Enhanced fMRI Visual Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CLR5OVL4}},
  note         = {Machine review of arXiv:2502.00412}
}
read the original abstract

fMRI (functional Magnetic Resonance Imaging) visual decoding involves decoding the original image from brain signals elicited by visual stimuli. This often relies on manually labeled ROIs (Regions of Interest) to select brain voxels. However, these ROIs can contain redundant information and noise, reducing decoding performance. Additionally, the lack of automated ROI labeling methods hinders the practical application of fMRI visual decoding technology, especially for new subjects. This work presents TROI (Trainable Region of Interest), a novel two-stage, data-driven ROI labeling method for cross-subject fMRI decoding tasks, particularly when subject samples are limited. TROI leverages labeled ROIs in the dataset to pretrain an image decoding backbone on a cross-subject dataset, enabling efficient optimization of the input layer for new subjects without retraining the entire model from scratch. In the first stage, we introduce a voxel selection method that combines sparse mask training and low-pass filtering to quickly generate the voxel mask and determine input layer dimensions. In the second stage, we apply a learning rate rewinding strategy to fine-tune the input layer for downstream tasks. Experimental results on the same small sample dataset as the baseline method for brain visual retrieval and reconstruction tasks show that our voxel selection method surpasses the state-of-the-art method MindEye2 with an annotated ROI mask.

Figures

Figures reproduced from arXiv: 2502.00412 by the authors.

Figure 1
Figure 1. Visual reconstructions from fMRI using baseline MindEye2 w/wo our proposed TROI in different training [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of TROI module structure followed by a residual MLP backbone and two task-specific submodules. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. TROI mask trained with the first 1h data of subject 1 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 13 canonical work pages

  1. [1]

    fmri-based decoding of visual information from human brain activity: A brief review

    Shuo Huang, Wei Shao, Mei-Ling Wang, and Dao-Qiang Zhang. fmri-based decoding of visual information from human brain activity: A brief review. International Journal of Automation and Computing, 18(2):170–184, 2021

  2. [2]

    Reconstructing the mind’s eye: fmri-to-image with contrastive learning and diffusion priors

    Paul Scotti, Atmadeep Banerjee, Jimmie Goode, Stepan Shabalin, Alex Nguyen, Aidan Dempster, Nathalie Verlinde, Elad Yundler, David Weisberg, Kenneth Norman, et al. Reconstructing the mind’s eye: fmri-to-image with contrastive learning and diffusion priors. Advances in Neural Information Processing Systems, 36, 2024

  3. [3]

    Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data

    Paul S Scotti, Mihir Tripathy, Cesar Kadir Torrico Villanueva, Reese Kneeland, Tong Chen, Ashutosh Narang, Charan Santhirasegaran, Jonathan Xu, Thomas Naselaris, Kenneth A Norman, et al. Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data. arXiv preprint arXiv:2403.11207, 2024

  4. [4]

    Reconstruction of perceived images from fmri patterns and semantic brain exploration using instance-conditioned gans

    Furkan Ozcelik, Bhavin Choksi, Milad Mozafari, Leila Reddy, and Rufin VanRullen. Reconstruction of perceived images from fmri patterns and semantic brain exploration using instance-conditioned gans. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2022

  5. [5]

    Natural scene reconstruction from fmri signals using generative latent diffusion

    Furkan Ozcelik and Rufin VanRullen. Natural scene reconstruction from fmri signals using generative latent diffusion. Scientific Reports, 13(1):15666, 2023

  6. [6]

    Reconstructing seen images from human brain activity via guided stochastic search

    Reese Kneeland, Jordyn Ojeda, Ghislain St-Yves, and Thomas Naselaris. Reconstructing seen images from human brain activity via guided stochastic search. ArXiv, 2023

  7. [7]

    Talairach and P

    J. Talairach and P. Tournoux. Co-planar Stereotaxic Atlas of the Human Brain: 3-dimensional Proportional System : an Approach to Cerebral Imaging. Thieme Publishers Series. G. Thieme, 1988

  8. [8]

    A probabilistic atlas and reference system for the human brain: International consortium for brain mapping (icbm)

    John Mazziotta, Arthur Toga, Alan Evans, Peter Fox, Jack Lancaster, Karl Zilles, Roger Woods, Tomas Paus, Gregory Simpson, Bruce Pike, et al. A probabilistic atlas and reference system for the human brain: International consortium for brain mapping (icbm). Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences, 356(1412):...

Show all 25 references
  1. [9]

    The wu-minn human connectome project: an overview

    David C Van Essen, Stephen M Smith, Deanna M Barch, Timothy EJ Behrens, Essa Yacoub, Kamil Ugurbil, Wu-Minn HCP Consortium, et al. The wu-minn human connectome project: an overview. Neuroimage, 80:62–79, 2013

  2. [10]

    The anatomical and functional organization of the human visual pulvinar

    Michael J Arcaro, Mark A Pinsk, and Sabine Kastner. The anatomical and functional organization of the human visual pulvinar. Journal of Neuroscience, 35(27):9848–9871, 2015

  3. [11]

    A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence

    Emily J Allen, Ghislain St-Yves, Yihan Wu, Jesse L Breedlove, Jacob S Prince, Logan T Dowdle, Matthias Nau, Brad Caron, Franco Pestilli, Ian Charest, et al. A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence. Nature neuroscience, 25(1):116–126, 2022

  4. [12]

    Investigations into resting-state connectivity using independent component analysis

    Christian F Beckmann, Marilena DeLuca, Joseph T Devlin, and Stephen M Smith. Investigations into resting-state connectivity using independent component analysis. Philosophical Transactions of the Royal Society B: Biological Sciences, 360(1457):1001–1013, 2005

  5. [13]

    The organization of the human cerebral cortex estimated by intrinsic functional connectivity

    BT Thomas Yeo, Fenna M Krienen, Jorge Sepulcre, Mert R Sabuncu, Danial Lashkari, Marisa Hollinshead, Joshua L Roffman, Jordan W Smoller, Lilla Zöllei, Jonathan R Polimeni, et al. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. Jour...

  6. [14]

    Regression shrinkage and selection via the lasso

    Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58(1):267–288, 1996

  7. [15]

    Mixco: Mix-up contrastive learning for visual representation

    Sungnyun Kim, Gihun Lee, Sangmin Bae, and Se-Young Yun. Mixco: Mix-up contrastive learning for visual representation. arXiv preprint arXiv:2010.06300, 2020

  8. [16]

    Over-training with mixup may hurt generalization

    Zixuan Liu, Ziqiao Wang, Hongyu Guo, and Yongyi Mao. Over-training with mixup may hurt generalization. arXiv preprint arXiv:2303.01475, 2023

  9. [17]

    Hierarchical text-conditional image generation with clip latents, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022

  10. [18]

    The human visual cortex

    Kalanit Grill-Spector and Rafael Malach. The human visual cortex. Annu. Rev. Neurosci., 27(1):649–677, 2004. 7

  11. [19]

    Channel pruning for accelerating very deep neural networks

    Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE international conference on computer vision, pages 1389–1397, 2017

  12. [20]

    Learning structured sparsity in deep neural networks

    Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. Advances in neural information processing systems, 29, 2016

  13. [21]

    Comparing rewinding and fine-tuning in neural network pruning

    Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. arXiv preprint arXiv:2003.02389, 2020

  14. [22]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings...

  15. [23]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004

  16. [24]

    Unsupervised learning of visual features by contrasting cluster assignments

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912–9924, 2020

  17. [25]

    Rethinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016. 8

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.