Pith. sign in

REVIEW 4 major objections 4 minor 74 references

High-Fidelity Functional Ultrasound Reconstruction via A Visual Auto-Regressive Framework

T0 review · 4 major / 4 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 UltraVAR, a visual auto-regressive framework, claims to synthesize functional ultrasound frames that preserve neurovascular coupling and vascular topology, and shows that augmenting training data with these frames raises rest-versus-task…

desk verdict First application of a visual auto-regressive model to fUS augmentation, but the downstream evaluation is confounded with class balancing and the significance claim has no statistical support. read the letter →

arxiv 2505.21530 v1 pith:BCOJGDE5 submitted 2025-05-23 eess.IV cs.AIcs.CV

Xuhang Chen , Zhuo Li , Yanyan Shen , Mufti Mahmud , Hieu Pham , Chi-Man Pun , Shuqiang Wang This is my paper
classification eess.IVcs.AIcs.CV
keywords functionalultrasoundimagingdataaugmentationvisualautoregressivemodelneurovascularcouplingimagereconstructiongenerativeAImedicalsynthesisdownstreamclassification
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

Functional ultrasound (fUS) imaging maps brain activity by tracking blood-volume changes, but ethical and technical constraints make fUS datasets small and skewed, which limits fair machine-learning models. This paper proposes UltraVAR, a generative augmentation framework that produces new fUS frames by predicting image tokens scale by scale, from coarse vascular layout to fine detail. The paper claims these synthetic frames keep the neurovascular coupling features and vascular topology that standard augmentations like flips or noise destroy, and that training classifiers on original-plus-UltraVAR data substantially improves rest-versus-task classification compared with six generative baselines. If the claim holds, generative augmentation could become a practical route to more diverse and fairer fUS datasets for brain-computer interfaces and neuromodulation research.

What carries the argument

The load-bearing mechanism is scale-by-scale auto-regressive token prediction, as in the pre-trained VAR model: the image is encoded by a VQVAE into discrete latent tokens, and a transformer predicts tokens for progressively finer patch sizes (1 to 32), with each scale conditioned on the previously generated coarser layout. On top of this, the Smooth Scaling Layer (SCL) applies a learned MLP to logits in windows of size 8 to enforce local coherence between neighboring tokens, and the Perception Enhancement Module (PEM) modulates decoder features through three cascaded Global Feature Modulation blocks, using a condition vector from the input, to suppress artifacts. Together they carry the claim by ensuring that generated frames retain the branching structure of microvessels and the subtle blood-volume differences that distinguish task from rest.

What would settle it

Train the same classifier on UltraVAR-augmented frames but replace task/rest labels with shuffled or permuted temporal blocks; if the accuracy gain persists, the signal is not neurovascular coupling but a static artifact. More directly, compute the cross-correlation between neural-activity traces and cerebral blood volume changes in synthetic frames and compare with real frames; near-zero correlation would falsify the coupling-preservation claim.

Watch

Extended reading notes

Core claim

The central discovery is that a next-scale visual auto-regressive generator, built on a VQVAE latent codebook, can reconstruct fUS frames at 128x128 that match the original data in SSIM (0.882/0.978), MS-SSIM (0.971/0.995), and FID (70.496/47.194), and that augmenting the human fUS training set with these samples lifts downstream classification F1 from 0.640 to 0.828 in Session1 and from 0.593 to 0.667 in Session2, outperforming CVAE, VQVAE, BigGAN, StyleGAN-XL, VQGAN, and DiT. The authors interpret the gain as evidence that the generated frames preserve the neurovascular coupling features (the rest-task contrast in blood-volume patterns) and the spatial topology of vascular networks, rather than merely looking similar. The two new modules, the Smooth Scaling Layer and the Perception Enhancement Module, are each shown by ablation to contribute to this outcome.

Load-bearing premise

The paper assumes that classifying individual 128x128 static frames as rest versus task is a valid proxy for preservation of neurovascular coupling, the dynamic temporal relationship between neural activity and microvascular hemodynamics; if the classification gain comes from a static confound like global intensity differences, the coupling claim is unsupported.

Editorial extensions

If this is right

  • Datasets augmented with UltraVAR improve rest-versus-task classification on held-out continuous blocks, with Session1 F1 rising from 0.640 to 0.828 and Session2 F1 from 0.593 to 0.667.
  • UltraVAR outperforms six generative baselines (CVAE, VQVAE, BigGAN, StyleGAN-XL, VQGAN, DiT) on SSIM, MS-SSIM, and FID for both sessions.
  • The Smooth Scaling Layer is the more critical component: removing it drops F1 to 0.675 (Session1) and 0.522 (Session2), while removing the Perception Enhancement Module drops it to 0.786 and 0.667.
  • Targeted augmentation of the minority class (Class1) with synthetic samples reduces class-imbalance bias and improves classification fairness metrics.
  • The authors propose the framework as a foundation for ultrasound-based neuromodulation and brain-computer interfaces, and as a candidate for adaptation to fetal ultrasound.

Reading between the lines

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

  • The validation is built entirely on static-frame classification, so the physiological claim about neurovascular coupling is only as strong as the assumption that frame-level differences carry the coupling signal; a direct temporal-dynamics test (e.g., hemodynamic response function shape in synthetic data) would be a natural next check.
  • The scale-by-scale prediction mechanism should transfer to other data-scarce physiological imaging modalities where spatial topology matters, such as fetal ultrasound or EEG-to-fNIRS synthesis, but the paper does not test this.
  • Because the authors deliberately avoided conventional augmentations while training UltraVAR, combining UltraVAR with classical geometric or noise augmentations may give further gains, though the paper does not report such a combination.
  • If the classification gains are driven by global intensity or other static features, the fairness improvement could be an artifact of rebalancing rather than physiological fidelity; a per-subject or per-block generalization test would clarify this.
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 manuscript proposes UltraVAR, a visual auto-regressive generative framework for functional ultrasound (fUS) image augmentation. The method combines a pretrained visual auto-regressive model with a VQVAE backbone and two proposed components—the Smooth Scaling Layer (SCL) and the Perception Enhancement Module (PEM)—to generate synthetic fUS frames that preserve vascular topology and, the authors claim, neurovascular coupling features. The framework is evaluated on a published human fUS dataset (Rabut et al.) using image-quality metrics (SSIM, MS-SSIM, FID), feature-space overlap (PCA), and a downstream rest/task classification task, with comparisons against six generative baselines. The central claims are that UltraVAR-generated data yield statistically significant improvements in downstream classification accuracy and that this improvement is attributable to preservation of neurovascular coupling.

Significance. The problem addressed—data scarcity in fUS imaging and the possible role of generative augmentation—is relevant and timely, and the proposed architecture is clearly motivated. The manuscript includes a pseudocode description of the training and inference procedure, which is useful for reproducibility, and the comparison against six strong generative baselines is a reasonable design. However, the current evidence for the central claims is undermined by a confounded evaluation design in which augmentation is conflated with class rebalancing, by an unsupported statistical-significance assertion, and by an internal inconsistency in the reported data-splitting protocol. If these issues are resolved through additional controlled experiments and proper statistical reporting, the contribution could be valuable for the medical imaging community.

major comments (4)
  1. [Section IV-D, Table I, Fig. 6] The downstream evaluation confounds the quality of UltraVAR-generated samples with simple class rebalancing. The fairness experiment adds 66 synthetic Class1 samples to a training set that originally contains 141 Class0 and 75 Class1 images; this changes the class ratio from approximately 65/35 to 50/50. The observed improvements in F1-score and Recall could therefore arise from reducing the class imbalance rather than from any property of the generated samples. A control condition is needed in which the same number of real Class1 samples is added (oversampling) or in which class weights are adjusted on the original data, so that the effect of UltraVAR fidelity can be isolated. Additionally, the comparisons against the six generative baselines in Table I do not specify the number or class composition of synthetic samples added for each baseline, so those comparisons are not controlled either.
  2. [Abstract and Conclusion (Section VI)] The abstract and conclusion assert that datasets augmented with UltraVAR yield 'statistically significant improvements' in downstream classification accuracy, but no statistical testing is reported anywhere in the manuscript. Table I and Fig. 6 present single-point estimates from what appears to be a single split, with no confidence intervals, no repeated-seed variance, and no paired significance tests. The authors should either report such statistics (e.g., multiple training runs with mean/std and a paired test) or temper the significance claim to a descriptive one.
  3. [Section IV-A vs Section IV-D] The description of the data split is internally inconsistent. Section IV-A states that holding out one task phase yields 720 training images and 270 test images for Session1 (and 1530 training images for Session2). Section IV-D, on the other hand, describes a random 80/20 split of 270 total images per session, yielding 216 training images (141 Class0, 75 Class1) and 54 test images. Both sections report identical original-data metrics for the downstream task (e.g., accuracy 0.800, F1 0.640 for Session1). These two protocols cannot both produce the same numbers, which raises serious questions about which protocol actually generated Table I and Fig. 6. The paper must clarify the precise experimental protocol and ensure all descriptions are consistent.
  4. [Section IV-D and Abstract] The central claim that UltraVAR 'preserves neurovascular coupling features' is not directly tested. Neurovascular coupling is a dynamic, temporal relationship between neural activity and microvascular hemodynamics, but the downstream validation in Section IV-D is static-frame classification of rest versus task on individual 128x128 images. A static classifier could exploit any static difference between rest and task blocks, such as global intensity or baseline vascular pattern, without necessarily reflecting the temporal coupling of neural and hemodynamic signals. To support the physiological-fidelity claim, the evaluation should include temporal metrics (e.g., comparison of event-related time courses, correlation of time-series statistics) or an explicitly temporally-aware downstream task. As it stands, the evidence is insufficient for the abstract's strong claim.
minor comments (4)
  1. [Algorithm 1] In line 8 of Algorithm 1, the update step uses an undefined variable L_total; this should be L_VQVAE. Additionally, the variable names 'idxs R' and 'scales BLC' in lines 17-18 are not defined or explained in the text.
  2. [Fig. 5 and Fig. 6] The axis label 'Precison' in both radar charts is a typo and should read 'Precision'.
  3. [General notation] The paper uses inconsistent spacing in the model name, mixing 'UltraV AR', 'UltraV AR', and 'UltraVAR' (e.g., in the abstract, Section I, and Fig. 1). Please standardize the notation.
  4. [Section IV-D] The term 'model fairness' is used to refer to improved performance on a class-imbalanced test set. This is an unconventional usage of 'fairness,' which typically refers to parity across demographic or protected groups. The authors should define what they mean by fairness or use a different term such as 'class-imbalance robustness.'

Circularity Check

1 steps flagged · score 6.0 of 10

Downstream validation confounds UltraVAR fidelity with minority-class oversampling, so the claimed physiological-preservation evidence partially reduces to a class-balancing effect.

  1. other [Section IV-D (Downstream Task Evaluation), Fig. 6 and Table I]
    "Recognizing the inherent class imbalance within the original data (Class1 being the minority), we augmented the training set with 66 additional synthetic Class1 samples... we randomly allocated 80% of the data for training (216 images: 141 Class0, 75 Class1) and reserved the remaining 20% for testing (54 images: 39 Class0, 15 Class1)."

    The experiment used to validate synthetic-image fidelity changes two variables at once: sample source and class balance. Adding exactly 66 synthetic Class1 samples converts the training set from 141 Class0 / 75 Class1 (roughly 1.88:1) to 141 Class0 / 141 Class1 (exactly 1:1). This is textbook minority oversampling, and any realistic Class1 samples - regardless of whether they preserve neurovascular coupling - would be expected to improve recall and F1 for the minority class. The paper attributes the F1 gain (Session1: 0.640 to 0.828; Session2: 0.593 to 0.667) to UltraVAR's preservation of physiological features, but no control adds 66 real Class1 samples, applies class weights, or undersamples Class0.

full rationale

No load-bearing self-citation chain was found: the core generative architecture is anchored to external foundations (VQVAE [46], VAR [29]) and the new SCL/PEM modules are defined through their own equations (Eq. 1-4) and losses (Eq. 5-7), which are not circular. The central vulnerability is the downstream validation loop. The paper's own setup in Section IV-D adds exactly 66 synthetic Class1 samples to a 141:75 imbalanced training set, making Class1 and Class0 equal at 141 each. This exactly reproduces a standard class-balancing intervention, so the observed F1 improvements are not uniquely attributable to UltraVAR's fidelity. The paper also asserts 'statistically significant improvements' without reporting confidence intervals, paired tests, or repeated-seed variance, and the Section IV-A protocol (held-out task phase, 720 training images) is inconsistent with the Section IV-D protocol (random 80/20 split of 270 images) despite identical 'Original' metrics; these issues reinforce that the headline comparison is not an isolated test of generative quality. Because the validation claim is partially self-referential but the generative method itself is independently defined, a score of 6 is appropriate.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim relies on a chain of proxy assumptions: static frames represent neurovascular coupling, classification accuracy measures physiological fidelity, and balanced augmentation is a fair test of generation quality. The paper does not directly measure coupling or temporal dynamics, and the fairness experiment is confounded by simultaneous class rebalancing.

free parameters (5)
  • Number of synthetic Class1 samples added = 66
    Chosen to counterbalance class imbalance in the fairness experiment; this choice, not generation quality, may drive the reported F1 gains.
  • VQVAE commitment loss weight beta = 0.25
    Set by hand in Eq. 6; standard, but still a selected hyperparameter that influences reconstruction fidelity.
  • Smooth Scaling Layer window size = 8
    Window size chosen for local logit coherence; no sensitivity analysis is reported.
  • PEM cascade depth = 3
    Three GFM modules are used; no ablation on depth or capacity is presented.
  • Patch size schedule = 1 to 32
    Multi-scale patch sizes ranging from 1 to 32 are selected without justification or sensitivity analysis.
assumptions (4)
  • domain assumption Temporal slicing of fUS time series into static 128x128 frames preserves the information needed to distinguish rest and task states.
    Section IV-A converts continuous streams into individual frames and uses them as training instances, discarding temporal dynamics central to neurovascular coupling.
  • domain assumption Classifier accuracy on static frames is a valid proxy for preservation of neurovascular coupling in generated images.
    Section IV-D uses downstream classification as the only quantitative evidence for physiological fidelity; no direct physiological or temporal validation is provided.
  • domain assumption Pre-trained natural-image VAR representations transfer to fUS after fine-tuning on a few hundred images.
    The framework is built on a pretrained VAR [29], but the paper does not analyze the domain gap or show performance before and after fine-tuning.
  • domain assumption Class labels from the block design are correct for every individual frame.
    Frames within rest and task blocks are labeled by block membership; boundary effects and delayed hemodynamic responses are not modeled.
invented entities (2)
  • Smooth Scaling Layer (SCL)
    purpose: Residual MLP on token logits intended to enforce local coherence between patch resolutions.
    Motivated by internal ablation only; no external or mechanistic validation that it preserves vascular topology.
  • Perception Enhancement Module (PEM)
    purpose: Three cascaded Global Feature Modulation blocks that modulate decoded features to suppress artifacts and enhance contrast.
    Benefit is shown only through the confounded downstream classification experiment; no independent perceptual or artifact evaluation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-Fidelity Functional Ultrasound Reconstruction via A Visual Auto-Regressive Framework." pith.science (2026). https://pith.science/paper/BCOJGDE5

@misc{pith2026250521530,
  author       = {Pith},
  title        = {Pith review of: High-Fidelity Functional Ultrasound Reconstruction via A Visual Auto-Regressive Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BCOJGDE5}},
  note         = {Machine review of arXiv:2505.21530}
}
read the original abstract

Functional ultrasound (fUS) imaging provides exceptional spatiotemporal resolution for neurovascular mapping, yet its practical application is significantly hampered by critical challenges. Foremost among these are data scarcity, arising from ethical considerations and signal degradation through the cranium, which collectively limit dataset diversity and compromise the fairness of downstream machine learning models.

Figures

Figures reproduced from arXiv: 2505.21530 by the authors.

Figure 1
Figure 1. Overview of the proposed UltraVAR framework. The architecture consists of a VQVAE backbone with encoder and decoder components, integrated with two modules: the Smooth Scaling Layer (SCL) and Perception Enhancement Module (PEM). The hierarchical reconstruction process refines image patches at progressive resolutions. A. Overview The proposed UltraVAR framework addresses the challenge of functional ultrasound image a… view at source ↗
Figure 2
Figure 2. Visual comparison of fUS images. This figure displays original fUS images alongside synthetic samples generated by UltraVAR and various baseline methods. Comparisons are shown for different brain activity classes (Class0 and Class1) across two experimental sessions (S1 stands for Session1 and S2 stands for Session2). UltraVARVQVAEVQGAN StyleGANXL DiT CVAEBIGGAN 0.75 0.80 0.85 0.90 0.882 0.868 0.811 0.786 0.747 0.744… view at source ↗
Figure 3
Figure 3. Quantitative comparison of image quality metrics across different generative models. The charts display SSIM, MS-SSIM, and FID values for Session1 (top row) and Session2 (bottom row). UltraVAR consistently outperforms other methods with higher SSIM and MS-SSIM values and lower FID scores. applied during the training of UltraVAR itself. This decision was made to preserve the intrinsic characteristics of the origi￾nal… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: PCA visualization comparing feature distributions of original fUS data, DiT-generated data, and UltraVAR-generated data. The closer clustering of UltraVAR samples to the original data, relative to DiT samples, indicates that UltraVAR generates data with representations…
Figure 5
Figure 5. Figure 5: Performance comparison on the downstream classification task. The radar charts illustrate classification metrics for different generative models on Session1 (left) and Session2 (right), with the proposed UltraVAR model showing superior performance. Accuracy Precision R…
Figure 6
Figure 6. Figure 6: Enhancement of downstream classification performance through UltraVAR data augmentation, indicating improved model fairness. The bar charts compare Accuracy, Precision, Recall, and F1-Score for models trained solely on the original limited dataset (w/o synthetic data) …
Figure 7
Figure 7. Figure 7: Ablation study results comparing classification performance. The radar charts show the impact of removing key components (PEM or SCL) from the proposed UltraVAR framework on Session1 (left) and Session2 (right), demonstrating the contribution of each module to overall …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 54 canonical work pages

  1. [1]

    UltraV AR, a visual auto-regressive framework is pro- posed for functional ultrasound image augmentation that preserves neurovascular coupling features

    work page

  2. [2]

    Smooth Scaling Layer is proposed to maintain infor- mation integrity during multi-scale feature propagation, ensuring coherent transitions between different patch resolutions in the hierarchical reconstruction process

    work page

  3. [3]

    Perception Enhancement Module is proposed to signifi- cantly reduce artifacts and enhance visual quality in the augmented fUS images. II. RELATEDWORK A. Functional Ultrasound Imaging Functional ultrasound imaging has rapidly emerged as a pivotal neuroimaging technique, distinguished by its excep- tional spatiotemporal resolution in monitoring cerebrovascu...

    work page

  4. [4]

    We employed the AdamW optimizer [55], an extension of the Adam optimizer that improves weight decay regularization, with hyperparameters set toβ 1 = 0.9and β2 = 0.999

    The number of epochs was chosen to ensure sufficient convergence of the model, while the batch size was determined by GPU memory constraints and the desire to have stable gradient estimates. We employed the AdamW optimizer [55], an extension of the Adam optimizer that improves weight decay regularization, with hyperparameters set toβ 1 = 0.9and β2 = 0.999...

    work page

  5. [5]

    Deep-fus: A deep learning platform for functional ultrasound imaging of the brain using sparse data,

    T. D. Ianni and R. D. Airan, “Deep-fus: A deep learning platform for functional ultrasound imaging of the brain using sparse data,”IEEE transactions on medical imaging, vol. 41, pp. 1813–1825, 2022

    work page 2022

  6. [6]

    M2trans: Multi-modal regularized coarse-to-fine transformer for ultra- sound image super-resolution,

    Z. Ni, R. Xiao, W. Yang, H. Wang, Z. Wang, L. Xiang, and L. Sun, “M2trans: Multi-modal regularized coarse-to-fine transformer for ultra- sound image super-resolution,”IEEE Journal of Biomedical and Health Informatics, pp. 1–12, 2024. doi: 10.1109/JBHI.2024.3454068

    arXiv 2024

  7. [7]

    Functional ultrasound localization microscopy reveals brain- wide neurovascular activity on a microscopic scale,

    N. Renaudin, C. Demen ´e, A. Dizeux, N. Ialy-Radio, S. Pezet, and M. Tanter, “Functional ultrasound localization microscopy reveals brain- wide neurovascular activity on a microscopic scale,”Nature methods, vol. 19, no. 8, pp. 1004–1012, 2022

    work page 2022

  8. [8]

    Ht-rcm: Hashimoto’s thyroiditis ultrasound image classification model based on res-fct and res-cam,

    W. Jiang, K. Chen, Z. Liang, T. Luo, G. Yue, Z. Zhao, W. Song, L. Zhao, and J. Wen, “Ht-rcm: Hashimoto’s thyroiditis ultrasound image classification model based on res-fct and res-cam,”IEEE Journal of Biomedical and Health Informatics, vol. 28, no. 2, pp. 941–951, 2024

    work page 2024

Show all 74 references
  1. [9]

    Taichinet: Negative-positive cross-attention network for breast lesion segmentation in ultrasound images,

    J. Wang, J. Liang, Y . Xiao, J. T. Zhou, Z. Fang, and F. Yang, “Taichinet: Negative-positive cross-attention network for breast lesion segmentation in ultrasound images,”IEEE Journal of Biomedical and Health Infor- matics, vol. 28, no. 3, pp. 1516–1527, 2024

    2024

  2. [10]

    Func- tional ultrasound imaging of human brain activity through an acous- tically transparent cranial window,

    C. Rabut, S. L. Norman, W. S. Griggs, J. J. Russin, K. Jann, V . Christopoulos, C. Liu, R. A. Andersen, and M. G. Shapiro, “Func- tional ultrasound imaging of human brain activity through an acous- tically transparent cranial window,”Science Translational Medicine, vol. 16, no...

    2024

  3. [11]

    Whole-brain functional ultrasound imaging in awake head- fixed mice,

    C. Brunner, M. Grillet, A. Urban, B. M. Roska, G. Montaldo, and E. Mac ´e, “Whole-brain functional ultrasound imaging in awake head- fixed mice,”Nature Protocols, vol. 16, pp. 3547–3571, 2021

    2021

  4. [12]

    Integrating genome-wide association and transcriptome prediction model identifies novel target genes for osteoporosis,

    M. Zhu, P. Yin, F. Hu, J. Jiang, L. Yin, Y . Li, and S. Wang, “Integrating genome-wide association and transcriptome prediction model identifies novel target genes for osteoporosis,”Osteoporosis International, vol. 32, no. 12, pp. 2493–2503, 2021

    2021

  5. [13]

    Prior-guided adver- sarial learning with hypergraph for predicting abnormal connections in alzheimer’s disease,

    Q. Zuo, H. Wu, C. P. Chen, B. Lei, and S. Wang, “Prior-guided adver- sarial learning with hypergraph for predicting abnormal connections in alzheimer’s disease,”IEEE Transactions on Cybernetics, vol. 54, no. 6, pp. 3652–3665, 2024

    2024

  6. [14]

    Bdht: Generative ai enables causality analysis for mild cognitive impairment,

    Q. Zuo, L. Chen, Y . Shen, M. Kwok-Po Ng, B. Lei, and S. Wang, “Bdht: Generative ai enables causality analysis for mild cognitive impairment,” IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 5601–5613, 2025

    2025

  7. [15]

    Eeg and functional ultrasound imaging in mobile rats,

    L.-A. Sieu, A. Bergel, E. Tiran, T. Deffieux, M. Pernot, J.-L. Gennisson, M. Tanter, and I. Cohen, “Eeg and functional ultrasound imaging in mobile rats,”Nature methods, vol. 12, no. 9, pp. 831–834, 2015

    2015

  8. [16]

    Neuroimaging retrieval via adaptive ensemble manifold learning for brain disease diagnosis,

    B. Lei, P. Yang, Y . Zhuo, F. Zhou, D. Ni, S. Chen, X. Xiao, and T. Wang, “Neuroimaging retrieval via adaptive ensemble manifold learning for brain disease diagnosis,”IEEE Journal of Biomedical and Health Informatics, vol. 23, no. 4, pp. 1661–1673, 2019

    2019

  9. [17]

    Adaptive cross-feature fusion network with inconsistency guidance for multi- modal brain tumor segmentation,

    G. Yue, G. Zhuo, T. Zhou, W. Liu, T. Wang, and Q. Jiang, “Adaptive cross-feature fusion network with inconsistency guidance for multi- modal brain tumor segmentation,”IEEE Journal of Biomedical and Health Informatics, pp. 1–11, 2023. doi: 10.1109/JBHI.2023.3347556

    2023

  10. [18]

    Single-trial decoding of movement intentions using functional ultrasound neuroimaging,

    S. L. Norman, D. Maresca, V . Christopoulos, W. S. Griggs, C. Demen ´e, M. Tanter, M. G. Shapiro, and R. A. Andersen, “Single-trial decoding of movement intentions using functional ultrasound neuroimaging,” Neuron, vol. 109, pp. 1554–1566.e4, 2020

    2020

  11. [19]

    Scdm: Unified representation learning for eeg-to-fnirs cross-modal generation in mi-bcis,

    Y . Li, Y . Wang, B. Lei, and S. Wang, “Scdm: Unified representation learning for eeg-to-fnirs cross-modal generation in mi-bcis,”IEEE Trans- actions on Medical Imaging, 2025. doi: 10.1109/TMI.2025.3532480

    2025

  12. [20]

    Catd: Unified representation learning for eeg-to-fmri cross-modal generation,

    W. Yao, Z. Lyu, M. Mahmud, N. Zhong, B. Lei, and S. Wang, “Catd: Unified representation learning for eeg-to-fmri cross-modal generation,”IEEE Transactions on Medical Imaging, 2025. doi: 10.1109/TMI.2025.3550206

    2025

  13. [21]

    Bidirectional mapping generative adversarial networks for brain mr to pet synthesis,

    S. Hu, B. Lei, S. Wang, Y . Wang, Z. Feng, and Y . Shen, “Bidirectional mapping generative adversarial networks for brain mr to pet synthesis,” IEEE Transactions on Medical Imaging, vol. 41, no. 1, pp. 145–157, 2021

    2021

  14. [22]

    Fine perceptive gans for brain mr image super- resolution in wavelet domain,

    S. You, B. Lei, S. Wang, C. K. Chui, A. C. Cheung, Y . Liu, M. Gan, G. Wu, and Y . Shen, “Fine perceptive gans for brain mr image super- resolution in wavelet domain,”IEEE transactions on neural networks and learning systems, vol. 34, no. 11, pp. 8802–8814, 2022

    2022

  15. [23]

    Generative ai enables eeg super-resolution via spatio-temporal adaptive diffusion learning,

    S. Wang, T. Zhou, Y . Shen, Y . Li, G. Huang, and Y . Hu, “Generative ai enables eeg super-resolution via spatio-temporal adaptive diffusion learning,”IEEE Transactions on Consumer Electronics, pp. 1–1, 2025. doi: 10.1109/TCE.2025.3528438

    2025

  16. [24]

    Progressive sub- band residual-learning network for mr image super resolution,

    X. Xue, Y . Wang, J. Li, Z. Jiao, Z. Ren, and X. Gao, “Progressive sub- band residual-learning network for mr image super resolution,”IEEE Journal of Biomedical and Health Informatics, vol. 24, no. 2, pp. 377– 386, 2020

    2020

  17. [25]

    Brain mr images super- resolution with the consistent features,

    S. You, Y . Shen, G. Wu, and S. Wang, “Brain mr images super- resolution with the consistent features,” inProceedings of the 2022 14th International Conference on Machine Learning and Computing, 2022, pp. 501–506

    2022

  18. [26]

    Addictive brain-network AUTHOR et al.: TITLE 13 identification by spatial attention recurrent network with feature selec- tion,

    C. Gong, X. Chen, B. Mughal, and S. Wang, “Addictive brain-network AUTHOR et al.: TITLE 13 identification by spatial attention recurrent network with feature selec- tion,”Brain Informatics, vol. 10, no. 1, p. 2, 2023

    2023

  19. [27]

    Addiction-related brain networks identification via graph diffusion reconstruction network,

    C. Jing, H. Kuai, H. Matsumoto, T. Yamaguchi, I. Y . Liao, and S. Wang, “Addiction-related brain networks identification via graph diffusion reconstruction network,”Brain Informatics, vol. 11, no. 1, p. 1, 2024

    2024

  20. [28]

    Enhanced generative adversarial network and target sample recognition method,

    S. Wang, S. Yanyan, and W. Zhang, “Enhanced generative adversarial network and target sample recognition method,” nov 2024, uS Patent 12,154,036

    2024

  21. [29]

    A new brain network construction paradigm for brain disorder via diffusion-based graph contrastive learning,

    Y . Zong, Q. Zuo, M. K.-P. Ng, B. Lei, and S. Wang, “A new brain network construction paradigm for brain disorder via diffusion-based graph contrastive learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. doi: 10.1109/TPAMI.2024.3442811

    2024

  22. [30]

    Es- timating addiction-related brain connectivity by prior-embedding graph generative adversarial networks,

    C. Jing, Y . Shen, S. Zhao, Y . Pan, C. P. Chen, B. Lei, and S. Wang, “Es- timating addiction-related brain connectivity by prior-embedding graph generative adversarial networks,”IEEE Transactions on Cybernetics, vol. 54, no. 9, pp. 5026–5039, 2024

    2024

  23. [31]

    Dec- gan: Decoupling generative adversarial network for detecting abnormal neural circuits in alzheimer’s disease,

    J. Pan, Q. Zuo, B. Wang, C. P. Chen, B. Lei, and S. Wang, “Dec- gan: Decoupling generative adversarial network for detecting abnormal neural circuits in alzheimer’s disease,”IEEE Transactions on Artificial Intelligence, vol. 5, no. 10, pp. 5050–5063, 2024

    2024

  24. [32]

    Adversarial learning based structural brain-network generative model for analyzing mild cognitive impairment,

    H. Kong, J. Pan, Y . Shen, and S. Wang, “Adversarial learning based structural brain-network generative model for analyzing mild cognitive impairment,” inChinese Conference on Pattern Recognition and Com- puter Vision (PRCV). Springer Nature Switzerland Cham, 2022, pp. 361–375

    2022

  25. [33]

    Functional ultrasound neu- roimaging,

    G. Montaldo, A. Urban, and E. Mac ´e, “Functional ultrasound neu- roimaging,”Annual Review of Neuroscience, vol. 45, no. 1, pp. 491–513, 2022

    2022

  26. [34]

    Functional ultrasound imaging of the brain,

    E. Mac ´e, G. Montaldo, I. Cohen, M. Baulac, M. Fink, and M. Tanter, “Functional ultrasound imaging of the brain,”Nature methods, vol. 8, no. 8, pp. 662–664, 2011

    2011

  27. [35]

    Functional ultrasound imaging of brain activity in human newborns,

    C. Demene, J. Baranger, M. Bernal, C. Delanoe, S. Auvin, V . Biran, M. Alison, J. Mairesse, E. Harribaud, M. Pernotet al., “Functional ultrasound imaging of brain activity in human newborns,”Science translational medicine, vol. 9, no. 411, p. eaah6756, 2017

    2017

  28. [36]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction,

    K. Tian, Y . Jiang, Z. Yuan, B. Peng, and L. Wang, “Visual autoregressive modeling: Scalable image generation via next-scale prediction,”arXiv preprint arXiv:2404.02905, 2024

    2024 arXiv

  29. [37]

    Towards fairness-aware and privacy-preserving enhanced collaborative learning for healthcare,

    F. Zhang, D. Zhai, G. Bai, J. Jiang, Q. Ye, X. Ji, and X. Liu, “Towards fairness-aware and privacy-preserving enhanced collaborative learning for healthcare,”Nature Communications, vol. 16, no. 1, p. 2852, 2025

    2025

  30. [38]

    Enhancing fairness in ai-enabled medical systems with the attribute neutral framework,

    L. Hu, D. Li, H. Liu, X. Chen, Y . Gao, S. Huang, X. Peng, X. Zhang, X. Bai, H. Yanget al., “Enhancing fairness in ai-enabled medical systems with the attribute neutral framework,”Nature Communications, vol. 15, no. 1, p. 8767, 2024

    2024

  31. [39]

    Achieving flexible fairness metrics in federated medical imaging,

    H. Xing, R. Sun, J. Ren, J. Wei, C.-M. Feng, X. Ding, Z. Guo, Y . Wang, Y . Hu, W. Weiet al., “Achieving flexible fairness metrics in federated medical imaging,”Nature Communications, vol. 16, no. 1, p. 3342, 2025

    2025

  32. [40]

    have explored skull-implant-compatible fUS configura- tions, thereby extending the potential for chronic neuroimag- ing in unconstrained human environments. Collectively, these pioneering developments chart the evolution of fUS from a specialized preclinical modality to an inc...

  33. [41]

    Star: Scale-wise text-to-image generation via auto-regressive representations,

    X. Ma, M. Zhou, T. Liang, Y . Bai, T. Zhao, H. Chen, and Y . Jin, “Star: Scale-wise text-to-image generation via auto-regressive representations,” arXiv preprint arXiv:2406.10797, 2024

    2024 arXiv

  34. [42]

    M-var: Decoupled scale-wise autoregressive modeling for high-quality image generation,

    S. Ren, Y . Yu, N. Ruiz, F. Wang, A. Yuille, and C. Xie, “M-var: Decoupled scale-wise autoregressive modeling for high-quality image generation,”arXiv preprint arXiv:2411.10433, 2024

    2024 arXiv

  35. [43]

    Bedside functional monitoring of the dynamic brain connectivity in human neonates,

    J. Baranger, C. Demene, A. Frerot, F. Faure, C. Delano ¨e, H. Serroune, A. Houdouin, J. Mairesse, V . Biran, O. Baudet al., “Bedside functional monitoring of the dynamic brain connectivity in human neonates,” Nature communications, vol. 12, no. 1, p. 1080, 2021

    2021

  36. [44]

    optimize latent space discretization, streamlining training and inference for large-scale datasets. These advancement in visual auto-regressive modeling not only push the boundaries of image reconstruction but also pro- vide valuable insights into the integration of generative...

  37. [45]

    Functional ultrasound (fus) during awake brain surgery: the clin- ical potential of intra-operative functional and vascular brain mapping,

    S. Soloukey, A. J. Vincent, D. D. Satoer, F. Mastik, M. Smits, C. M. Dirven, C. Strydis, J. G. Bosch, A. F. van der Steen, C. I. De Zeeuw et al., “Functional ultrasound (fus) during awake brain surgery: the clin- ical potential of intra-operative functional and vascular brain ...

    2020

  38. [46]

    Functional ultrasound imaging of the human spinal cord,

    K. A. Agyeman, D. J. Lee, J. Russin, E. I. Kreydin, W. Choi, A. Abedi, Y . T. Lo, J. Cavaleri, K. Wu, V . R. Edgertonet al., “Functional ultrasound imaging of the human spinal cord,”Neuron, vol. 112, no. 10, pp. 1710– 1722, 2024

    2024

  39. [47]

    Conditional reconstruction is achieved by in- corporating class embeddings that guide the reconstruction process

    with rotary positional embeddings to process the token representations. Conditional reconstruction is achieved by in- corporating class embeddings that guide the reconstruction process. We employ classifier-free guidance during inference to strengthen the conditioning signal, ...

  40. [48]

    The emergence of functional ultrasound for noninvasive brain–computer interface,

    H. Zheng, L. Niu, W. Qiu, D. Liang, X. Long, G. Li, Z. Liu, and L. Meng, “The emergence of functional ultrasound for noninvasive brain–computer interface,”Research, vol. 6, p. 0200, 2023

    2023

  41. [49]

    P09. 02. aa window to the brain: Monitoring the human brain using functional ultrasound (fus)-imaging in freely moving subjects with an acoustically favorable skull implant,

    S. Soloukey, L. Verhoef, F. Mastik, M. Brown, B. Generowicz, D. Satoer, C. Dirven, M. Smits, S. Koekkoek, A. Vincentet al., “P09. 02. aa window to the brain: Monitoring the human brain using functional ultrasound (fus)-imaging in freely moving subjects with an acoustically fav...

    2024

  42. [50]

    Imagefolder: Autoregressive image generation with folded tokens,

    X. Li, K. Qiu, H. Chen, J. Kuen, J. Gu, B. Raj, and Z. Lin, “Imagefolder: Autoregressive image generation with folded tokens,”arXiv preprint arXiv:2410.01756, 2024

    2024 arXiv

  43. [51]

    Xq- gan: An open-source image tokenization framework for autoregressive generation,

    X. Li, K. Qiu, H. Chen, J. Kuen, J. Gu, J. Wang, Z. Lin, and B. Raj, “Xq- gan: An open-source image tokenization framework for autoregressive generation,”arXiv preprint arXiv:2412.01762, 2024

    2024 arXiv

  44. [52]

    Conditional diffusion model-based data augmentation for alzheimer’s prediction,

    W. Yao, Y . Shen, F. Nicolls, and S.-Q. Wang, “Conditional diffusion model-based data augmentation for alzheimer’s prediction,” inInter- national Conference on Neural Computing for Advanced Applications. Springer Nature Singapore Singapore, 2023, pp. 33–46

    2023

  45. [53]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” inNeural Information Processing Systems, 2017

    2017

  46. [54]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in neural information processing systems, 2017, pp. 5998–6008

    2017

  47. [55]

    Uwformer: Underwater image enhancement via a semi-supervised multi-scale trans- former,

    W. Chen, Y . Lei, S. Luo, Z. Zhou, M. Li, and C.-M. Pun, “Uwformer: Underwater image enhancement via a semi-supervised multi-scale trans- former,” in2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024, pp. 1–8

    2024

  48. [56]

    Underwater image restoration through a prior guided hybrid sense approach and extensive benchmark analysis,

    X. Guo, X. Chen, S. Wang, and C.-M. Pun, “Underwater image restoration through a prior guided hybrid sense approach and extensive benchmark analysis,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  49. [57]

    Learning structured output representa- tion using deep conditional generative models,

    K. Sohn, H. Lee, and X. Yan, “Learning structured output representa- tion using deep conditional generative models,” inNeural Information Processing Systems, 2015

    2015

  50. [58]

    Large scale GAN training for high fidelity natural image synthesis,

    A. Brock, J. Donahue, and K. Simonyan, “Large scale GAN training for high fidelity natural image synthesis,” inInternational Conference on Learning Representations, 2019

    2019

  51. [59]

    Stylegan-xl: Scaling stylegan to large diverse datasets,

    A. Sauer, K. Schwarz, and A. Geiger, “Stylegan-xl: Scaling stylegan to large diverse datasets,” inACM SIGGRAPH 2022 conference proceed- ings, 2022, pp. 1–10

    2022

  52. [60]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12 873–12 883

    2021

  53. [61]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4195–4205

    2023

  54. [62]

    Fixing weight decay regularization in adam,

    I. Loshchilov, F. Hutteret al., “Fixing weight decay regularization in adam,”arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  55. [63]

    SGDR: stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “SGDR: stochastic gradient descent with warm restarts,” inProceedings of International Conference on Learning Representations, 2017

    2017

  56. [64]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  57. [65]

    Multiscale structural similarity for image quality assessment,

    Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multiscale structural similarity for image quality assessment,” inThe Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, vol. 2, 2003, pp. 1398–1402

    2003

  58. [66]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,”Advances in neural information processing systems, vol. 30, 2017

    2017

  59. [67]

    Flexible brain–computer interfaces,

    X. Tang, H. Shen, S. Zhao, N. Li, and J. Liu, “Flexible brain–computer interfaces,”Nature Electronics, vol. 6, no. 2, pp. 109–118, 2023

    2023

  60. [68]

    Interface, interaction, and in- telligence in generalized brain–computer interfaces,

    X. Gao, Y . Wang, X. Chen, and S. Gao, “Interface, interaction, and in- telligence in generalized brain–computer interfaces,”Trends in cognitive sciences, vol. 25, no. 8, pp. 671–684, 2021

    2021

  61. [69]

    The current re- search of combining multi-modal brain-computer interfaces with virtual reality,

    D. Wen, B. Liang, Y . Zhou, H. Chen, and T.-P. Jung, “The current re- search of combining multi-modal brain-computer interfaces with virtual reality,”IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 9, pp. 3278–3287, 2021

    2021

  62. [70]

    Neuromodulation of brain states,

    S.-H. Lee and Y . Dan, “Neuromodulation of brain states,”neuron, vol. 76, no. 1, pp. 209–222, 2012

    2012

  63. [71]

    Neuromodulation for chronic pain,

    H. Knotkova, C. Hamani, E. Sivanesan, M. F. E. Le Beuffe, J. Y . Moon, S. P. Cohen, and M. A. Huntoon, “Neuromodulation for chronic pain,” The Lancet, vol. 397, no. 10289, pp. 2111–2124, 2021

    2021

  64. [72]

    Neuro- modulation in epilepsy: state-of-the-art approved therapies,

    P. Ryvlin, S. Rheims, L. J. Hirsch, A. Sokolov, and L. Jehi, “Neuro- modulation in epilepsy: state-of-the-art approved therapies,”The Lancet Neurology, vol. 20, no. 12, pp. 1038–1047, 2021

    2021

  65. [73]

    Multi-modal mood reader: Pre-trained model empowers cross-subject emotion recognition,

    Y . Dong, X. Chen, Y . Shen, M. K.-P. Ng, T. Qian, and S. Wang, “Multi-modal mood reader: Pre-trained model empowers cross-subject emotion recognition,” inInternational Conference on Neural Computing for Advanced Applications. Springer Nature Singapore Singapore, 2024, pp. 178–192

    2024

  66. [74]

    Smart diagnosis assistance method to solve results of inaccurate classification of image, and terminal based on medical images,

    S. Wang, W. Yu, Z. Chenet al., “Smart diagnosis assistance method to solve results of inaccurate classification of image, and terminal based on medical images,” mar 2025, uS Patent 12,254,684

    2025

Pith tools

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