Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Parameter-Efficient Fine-Tuning of 3D DDPM for MRI Image Generation Using Tensor Networks

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TenVOO fine-tunes a 3D U-Net diffusion model for brain MRI by encoding convolution updates as tensor networks, matching low-rank baselines on structural fidelity with only about 0.3% of the original trainable parameters.

desk verdict TenVOO is a plausible tensor-network PEFT extension to 3D convolutions with open code, but the SOTA claim rests on an underspecified MS-SSIM evaluation and needs major revision before the headline can be trusted. read the letter →

arxiv 2507.18112 v1 pith:5BAE7UPR submitted 2025-07-24 eess.IV cs.AIcs.CV

Binghua Li , Ziqing Chang , Tong Liang , Chao Li , Toshihisa Tanaka , Shigeki Aoki , Qibin Zhao , Zhe Sun This is my paper
classification eess.IVcs.AIcs.CV
keywords tensornetworksparameter-efficientfine-tuningdenoisingdiffusionprobabilisticmodels3DconvolutionMRIimagegenerationU-Netlow-rankadaptationmedicalimaging
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

The paper proposes TenVOO, a parameter-efficient fine-tuning scheme for 3D U-Net-based denoising diffusion models used to generate brain MRI. The core idea is to represent the update to each 3D convolution kernel as a tensor network—a small set of lower-dimensional tensors joined by contractions—so that spatial dependencies are captured with far fewer trainable parameters than full fine-tuning. The paper claims that, fine-tuning a model pretrained on tens of thousands of T1-weighted brain scans, TenVOO reaches the highest multi-scale structural similarity (MS-SSIM) among the tested parameter-efficient baselines on three downstream datasets while training only about 0.3% of the original model's parameters. The payoff, if true, is that one large pretrained MRI generator can be adapted to many clinical datasets cheaply, without storing a full model copy per task.

What carries the argument

The load-bearing object is the tensor network $T_\theta$ in Eq. (5): a graph of small core tensors whose contractions produce the full 3D convolution update tensor $\Delta W$. TenVOO-L factorizes the input and output channel dimensions into products $C_{\mathrm{in}} = i_1 i_2 i_3$ and $C_{\mathrm{out}} = o_1 o_2 o_3$, then strands the spatial dimensions $k_d, k_h, k_w$ as separate legs so spatial information is carried by contraction rather than by a dense low-rank factor. TenVOO-Q instead assigns spatial dimensions to different cores in a QuanTA-style circuit. The rank $r$ sets the parameter count and, the ablation suggests, the representational capacity; the paper fixes $r=4$ for the main experiments. Initialization keeps the update at zero by subtracting a frozen copy of the initial tensor network, following QuanTA's routine.

What would settle it

Train TenVOO on a synthetic 3D convolution task where the correct update is a known rank-8 kernel, such as a depth shift combined with channel mixing. If the model's loss or structural similarity plateaus below an unconstrained fine-tune at matched data and budget, the rank-4 constraint, not the tensor-network idea, is the active limit.

Watch

Extended reading notes

Core claim

The central claim is that a tensor-network parameterization of 3D convolution weight updates—called the Tensor Volumetric Operator—can serve as a parameter-efficient fine-tuning mechanism for U-Net-based DDPMs. For each 3D convolutional layer, TenVOO writes the update as $\Delta W = \mathrm{reshape}(T_\theta(t_1,\ldots,t_k), [C_{\mathrm{out}}, C_{\mathrm{in}}, k_d, k_h, k_w])$, where $T_\theta$ is a tensor network built from small core tensors. Two variants are given: TenVOO-L, which tensorizes the channel dimensions and keeps spatial dimensions separate, and TenVOO-Q, an extension of the QuanTA structure that assigns spatial dimensions to different cores. Fine-tuning updates only the core tensors. The paper reports that on three downstream brain MRI datasets—an Alzheimer's cohort, a Parkinson's cohort, and a brain-tumor cohort—TenVOO matches or improves FID/MMD and achieves the highest MS-SSIM among the tested PEFT methods, with 0.58–0.60 million trainable parameters versus 166.67 million for full fine-tuning.

Load-bearing premise

The load-bearing premise is that the chosen tensor-network layout with rank 4 can express the 3D-convolution adjustments each dataset needs; the paper does not prove this, and its own ablation shows structural similarity still rising at higher ranks.

Editorial extensions

If this is right

  • One pretrained 3D brain-MRI diffusion model can be reused across many clinical datasets by storing only small tensor cores per task, cutting per-task storage from hundreds of millions of parameters to under a million.
  • Tensor-network adapters can express spatial structure that plain low-rank LoRA misses, making them a natural fit for convolutional backbones where low-rank updates are too rigid.
  • Because the update is an additive correction, the fine-tuned weights can be merged back into the pretrained model at inference, so the adapted model runs at the same speed as the original.
  • The same decomposition applies to any 3D convolutional layer, so the method generalizes beyond MRI to other 3D generative and medical-imaging models.

Reading between the lines

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

  • Beyond the paper, the rank-4 ceiling is the most likely failure point: the paper's own rank ablation shows MS-SSIM still climbing at rank 6, so on tasks demanding larger convolution shifts a rank-4 network may underperform.
  • Beyond the paper, TenVOO's largest gains appear on the dataset most distant from the pretraining distribution, suggesting tensor-network adapters help most when adaptation requires structural rather than intensity-level changes.
  • Beyond the paper, because the decomposition places each spatial dimension on its own leg, the same construction extends to 4D data or to latent-space generators, with parameter growth roughly linear in the number of added dimensions at fixed rank.
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

5 major / 5 minor

Summary. The paper proposes TenVOO, a parameter-efficient fine-tuning (PEFT) method for 3D U-Net-based denoising diffusion probabilistic models (DDPMs) in MRI image generation. TenVOO represents each 3D convolution weight update as a tensor network (TenVOO-L and TenVOO-Q variants), with an initialization scheme that keeps the initial effective update at zero, and it also applies QuanTA-style updates to linear layers. The method is evaluated by fine-tuning a DDPM pretrained on 59,830 UK Biobank scans on three downstream datasets (ADNI, PPMI, BraTS2021), comparing FID, MMD, and MS-SSIM against LoRA, LoKr, LoHa, and full fine-tuning. The paper claims state-of-the-art MS-SSIM performance with only about 0.3% of the original model's trainable parameters, and the ablation study examines the effect of tensor rank on parameter count and MS-SSIM.

Significance. If the empirical claims are substantiated, TenVOO would be a useful demonstration that tensor-network-structured updates can capture spatial dependencies in 3D medical image generation more effectively than low-rank baselines at a lower parameter count. The paper is, to my knowledge, the first to apply tensor-network PEFT to 3D DDPMs for MRI, and it ships a public code link, which is a concrete strength. However, the evidence as presented is mixed: the method does not consistently outperform baselines on FID/MMD, its MS-SSIM advantage is not accompanied by a specification of how the metric is computed in an unconditional generation setting, and there are no error bars or multiple-seed results. The central claim is therefore defensible but not yet established.

major comments (5)
  1. [Section 3.1, Table 1] The MS-SSIM evaluation protocol is not specified. MS-SSIM is a full-reference metric and the paper never states how generated volumes are paired with real volumes when the DDPM is unconditional. Without this pairing rule, the reported MS-SSIM values cannot be interpreted, and the observed profile (e.g., TenVOO-L on ADNI and PPMI having worse FID but better MS-SSIM than LoRA) is exactly what an over-smoothed 'average brain' output would produce. Please specify the pairing rule (e.g., random pairing, nearest-neighbor in some embedding, or per-real-volume matching) and, ideally, include a smoothing baseline to show that the MS-SSIM gain is not simply a smoothness artifact.
  2. [Tables 1 and 2] There are no error bars or multiple-seed runs, so the reported improvements cannot be distinguished from training noise. For example, TenVOO-L's ADNI MS-SSIM of 0.663 is below Full-FT's 0.701, TenVOO-Q's PPMI FID of 19.585 is worse than LoRA's 14.609, and the BraTS2021 MS-SSIM differences are large but based on a single run. Please report means and standard deviations over at least three seeds, or otherwise justify why single-run numbers support the stated conclusions.
  3. [Section 3.2, Figure 4(b)] The rank ablation does not support the choice of rank 4. Figure 4(b) shows MS-SSIM continuing to increase as rank goes from 4 to 6, with no plateau, so the reported configuration is a capacity ceiling rather than a converged optimum. The text says 'We set tensor rank as 4 and constrain our TenVOO to have the least number of trainable parameters,' but this conflates a chosen hyperparameter with an optimized one. Please either justify rank 4 as a deliberate trade-off with evidence that higher ranks do not change the conclusions, or report results at a rank where performance saturates.
  4. [Abstract and Section 3.2] The claim of 'state-of-the-art performance in MS-SSIM' is overstated. In Table 1, Full-FT achieves MS-SSIM 0.701 on ADNI, higher than TenVOO-L's 0.663, and TenVOO-Q's ADNI MS-SSIM of 0.504 is below LoRA's 0.511. The statement is only defensible if restricted to 'state-of-the-art among PEFT baselines,' and even then it must be reconciled with the fact that FID and MMD are not consistently favorable (e.g., TenVOO-L ADNI FID 17.349 vs LoRA 16.794; TenVOO-L PPMI FID 16.605 vs LoRA 14.609). Please qualify the claim and discuss the metric-dependent nature of the results.
  5. [Section 3.1] The BraTS2021 evaluation uses only 327 of 1,251 training samples after filtering 'based on visual quality,' following reference [11]. This subjective filter could bias results, and BraTS2021 is precisely the dataset where TenVOO shows its largest improvements. Please specify the exact filtering criteria, report how many samples were excluded for each reason, and ideally show that the main conclusions are robust to including the unfiltered set or to using a principled exclusion rule.
minor comments (5)
  1. [Section 3.2] The sentence 'Notably, we observe that TenVOOs yield much lower FID and MMD scores on PPMI' contradicts Table 1, where TenVOO-L has PPMI FID 16.605 versus LoRA's 14.609 and TenVOO-Q has FID 19.585. Please correct this statement.
  2. [Equation (6)] The notation 'r2', 'r3', 'r4' in Eq. (6) should be written as r^2, r^3, r^4 to avoid ambiguity, and the parameter-count expressions would benefit from a derivation or a worked example for one layer.
  3. [Section 2.1] There are minor typos in the text, including 'backword process' and 'covolution'; please proofread for these and similar errors.
  4. [Section 2.4] The initialization in Eq. (8) is clear, but the sentence 'Here only core tensors {t_i^(0)} in ∆W are updated' is confusing because ∆W is defined as the full TN and the frozen copy is Tbar. Please clarify that t_i are the trainable cores and Tbar is frozen.
  5. [Figure 2] The caption uses inconsistent capitalization ('TenVoo-L' vs 'TenVOO-L'); also, panel (b) is referred to as 'LoRA 3D' but the text calls it 'LoCon' from Lycoris; please make the terminology consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: TenVOO's claims are parameterization definitions plus externally benchmarked empirical results.

full rationale

The paper's derivation chain consists of parameterization definitions (Eqs. 5-8) followed by empirical comparisons. There is no fitted parameter that is later relabeled as a prediction; the tensor rank r=4 is a hyperparameter, and the ablation in Fig. 4b explicitly varies it, which is the opposite of fitting it to the target metric. The only external load-bearing ingredient is QuanTA's initialization, and that is cited to external prior work rather than to this paper's own results. Numeric claims (Table 1) are held-out empirical benchmarks against independent baselines. Concerns about the unspecified MS-SSIM pairing protocol in unconditional generation are evaluation validity issues, not circularity; they affect how much weight the SOTA claim carries but do not make the derivation reduce to its inputs.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The paper's central claim rests on the modeling premise that a rank-4 tensor network can represent the needed 3D convolution updates. The only numbers introduced are experiment hyperparameters and the BraTS2021 filter; no physical or architectural entities are postulated beyond the TenVOO parameterization itself, which is the method under test.

free parameters (3)
  • Tensor rank r = 4
    Chosen to minimize the number of trainable parameters. Ablation (Fig. 4) shows MS-SSIM increases with rank, so r=4 is a hand-picked efficiency floor rather than an optimized value.
  • Learning rate = 0.00005
    Fixed optimizer hyperparameter across all runs; no sensitivity analysis is reported.
  • BraTS2021 visual quality filter = 327 of 1,251 training scans retained
    Data subset chosen based on visual quality, following [11]. This hand-selected subset affects the evaluation and likely inflates performance metrics.
assumptions (6)
  • standard math DDPM forward and reverse process equations (1)-(2) are adopted without modification.
    The method relies on standard denoising diffusion probabilistic model theory; no changes are made.
  • domain assumption A TN-parameterized ΔW can be reshaped into a 3D convolution kernel and the sum W+ΔW captures the needed adaptation.
    Section 2.2, eq. (5). This is the core modeling assumption, supported only by the paper's experiments.
  • ad hoc to paper The chosen initialization with a frozen copy of the TN maintains stable fine-tuning.
    Section 2.4, eq. (8). Empirically motivated, with no theoretical guarantee.
  • domain assumption MS-SSIM between generated and reference scans is a valid proxy for capturing spatial dependencies.
    Section 3.1 uses MS-SSIM as the main evidence for the method's benefit; no validation of this proxy is given.
  • domain assumption Med3D embeddings provide a meaningful feature space for FID and MMD in these MRI datasets.
    Section 3.1 follows prior work [11,27] but does not verify the encoder's suitability.
  • domain assumption All datasets are preprocessed to the same MNI152 space and resolution, making voxel-level comparisons valid.
    Section 3.1 describes resampling, skull stripping, and registration; if alignments differ, MS-SSIM comparisons are less meaningful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Parameter-Efficient Fine-Tuning of 3D DDPM for MRI Image Generation Using Tensor Networks." pith.science (2026). https://pith.science/paper/5BAE7UPR

@misc{pith2026250718112,
  author       = {Pith},
  title        = {Pith review of: Parameter-Efficient Fine-Tuning of 3D DDPM for MRI Image Generation Using Tensor Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5BAE7UPR}},
  note         = {Machine review of arXiv:2507.18112}
}
read the original abstract

We address the challenge of parameter-efficient fine-tuning (PEFT) for three-dimensional (3D) U-Net-based denoising diffusion probabilistic models (DDPMs) in magnetic resonance imaging (MRI) image generation. Despite its practical significance, research on parameter-efficient representations of 3D convolution operations remains limited. To bridge this gap, we propose Tensor Volumetric Operator (TenVOO), a novel PEFT method specifically designed for fine-tuning DDPMs with 3D convolutional backbones. Leveraging tensor network modeling, TenVOO represents 3D convolution kernels with lower-dimensional tensors, effectively capturing complex spatial dependencies during fine-tuning with few parameters. We evaluate TenVOO on three downstream brain MRI datasets-ADNI, PPMI, and BraTS2021-by fine-tuning a DDPM pretrained on 59,830 T1-weighted brain MRI scans from the UK Biobank. Our results demonstrate that TenVOO achieves state-of-the-art performance in multi-scale structural similarity index measure (MS-SSIM), outperforming existing approaches in capturing spatial dependencies while requiring only 0.3% of the trainable parameters of the original model. Our code is available at: https://github.com/xiaovhua/tenvoo

Figures

Figures reproduced from arXiv: 2507.18112 by the authors.

Figure 1
Figure 1. Illustration of the U-Net model used for our DDPM and the TenVOO framework for 3D convolutional layers. TenVOO optimizes only the weight updates ∆W via a lightweight tensor network during fine-tuning, while preserving the intricate spatial dependencies of convolution kernels throught tensor contraction. effectively capture complex data distributions. However, their large-scale archi￾tecture poses significant computa… view at source ↗
Figure 2
Figure 2. Structure of (a) LoRA for 2D convolution1 ; (b) LoRA for 3D covolution2 ; (c) TenVOO-L for 3D convolution and (d) TenVOO-Q for 3D convolution. The thick yellow connections represent the tunable rank of the TN. where ∆W is the low-rank update of the 2D convolutional weights, r corresponds to the LoRA rank, and Cout, Cin, kh, kw represent the number of output channels, input channels, and the kernel size in height and… view at source ↗
Figure 3
Figure 3. Visualization results. For each image pair, the top scan is generated by the jointly fine-tuned model, while the bottom scan is generated by the adapted model. （a） 0.37M 0.41M 0.40M 0.60M 0.58M 1.05M 1.04M 0.38M （b） [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Results of the ablation study on rank. (a) illustrates how the number of param￾eters changes as the rank increases. (b) demonstrates the impact of rank on MS-SSIM performance. in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. GUST: Quantifying Free-Form Geometric Uncertainty of Metamaterials Using Small Data

    cs.LG 2025-05 conditional novelty 5.0 of 10

    GUST combines synthetic-data pretraining with transfer learning on a conditional diffusion model to quantify free-form geometric uncertainty in manufactured metamaterials from small real-world datasets.

Reference graph

Works this paper leans on

34 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [11]

    IEEE Journal of Biomedical and Health Informatics (2024)

    Dorjsembe, Z., Pao, H.K., Odonchimed, S., Xiao, F.: Conditional diffusion models for semantic 3D brain MRI synthesis. IEEE Journal of Biomedical and Health Informatics (2024)

    work page 2024

  2. [1]

    In: AICS

    Ali, H., Murad, S., Shah, Z.: Spot the fake lungs: Generating synthetic medical images using neural diffusion models. In: AICS. pp. 32–39. Springer (2022)

    work page 2022

  3. [2]

    arXiv preprint arXiv:2107.02314 (2021)

    Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Fara- hani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The RSNA-ASNR- MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)

    arXiv 2021

  4. [3]

    arXiv preprint arXiv:2402.01376 (2024)

    Bershatsky, D., Cherniuk, D., Daulbaev, T., Mikhalev, A., Oseledets, I.: LoTR: Low tensor rank weight adaptation. arXiv preprint arXiv:2402.01376 (2024)

    arXiv 2024

  5. [4]

    Brain127(4), 791– 800 (2004)

    Burton, E.J., McKeith, I.G., Burn, D.J., Williams, E.D., O’Brien, J.T.: Cerebral atrophy in Parkinson’s disease with and without dementia: a comparison with Alzheimer’s disease, dementia with lewy bodies and controls. Brain127(4), 791– 800 (2004)

    work page 2004

  6. [5]

    arXiv preprint arXiv:2211.02701 (2022)

    Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)

    arXiv 2022

  7. [6]

    arXiv preprint arXiv:2306.07967 (2023)

    Chavan, A., Liu, Z., Gupta, D., Xing, E., Shen, Z.: One-for-all: Generalized LoRA for parameter-efficient fine-tuning. arXiv preprint arXiv:2306.07967 (2023)

    arXiv 2023

  8. [7]

    arXiv preprint arXiv:1904.00625 (2019)

    Chen, S., Ma, K., Zheng, Y.: Med3D: Transfer learning for 3D medical image analysis. arXiv preprint arXiv:1904.00625 (2019)

    arXiv 2019

Show all 34 references
  1. [8]

    In: NeurIPS (2024),https://openreview.net/forum?id=EfpZNpkrm2

    Chen, Z., Dangovski, R., Loh, C., Dugan, O.M., Luo, D., Soljacic, M.: QuanTA: Efficient high-rank fine-tuning of LLMs with quantum-informed tensor adaptation. In: NeurIPS (2024),https://openreview.net/forum?id=EfpZNpkrm2

    2024

  2. [9]

    IEEE Transactions on Medical Imaging42(4), 922– 934 (2022)

    Chung, H., Lee, E.S., Ye, J.C.: MR image denoising and super-resolution using regularized reverse diffusion. IEEE Transactions on Medical Imaging42(4), 922– 934 (2022)

    2022

  3. [10]

    arXiv preprint arXiv:1407.3124 (2014)

    Cichocki, A.: Tensor networks for big data analytics and large-scale optimization problems. arXiv preprint arXiv:1407.3124 (2014)

    2014 arXiv

  4. [12]

    arXiv preprint arXiv:2212.10650 (2022) 10 Binghua Li, Ziqing Chang, Tong Liang et al

    Edalati, A., Tahaei, M., Kobyzev, I., Nia, V.P., Clark, J.J., Rezagholizadeh, M.: Krona: Parameter efficient tuning with Kronecker adapter. arXiv preprint arXiv:2212.10650 (2022) 10 Binghua Li, Ziqing Chang, Tong Liang et al

    2022 arXiv

  5. [13]

    Gigascience 13, giae049 (2024)

    Gaser, C., Dahnke, R., Thompson, P.M., Kurth, F., Luders, E., Initiative, A.D.N., et al.: CAT: a computational anatomy toolbox for the analysis of structural MRI data. Gigascience 13, giae049 (2024)

    2024

  6. [14]

    The Journal of Machine Learning Research13(1), 723–773 (2012)

    Gretton, A., Borgwardt, K.M., Rasch, M.J., Schölkopf, B., Smola, A.: A kernel two-sample test. The Journal of Machine Learning Research13(1), 723–773 (2012)

    2012

  7. [15]

    In: ISBI

    Guo, X., Yang, Y., Ye, C., Lu, S., Peng, B., Huang, H., Xiang, Y., Ma, T.: Acceler- ating diffusion models via pre-segmentation diffusion sampling for medical image segmentation. In: ISBI. pp. 1–5. IEEE (2023)

    2023

  8. [16]

    In: NeurIPS (2017)

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In: NeurIPS (2017)

    2017

  9. [17]

    In: NeurIPS

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS. pp. 6840–6851 (2020)

    2020

  10. [18]

    In: ICML

    Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Ges- mundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for NLP. In: ICML. pp. 2790–2799. PMLR (2019)

    2019

  11. [19]

    In: ICLR (2022),https: //openreview.net/forum?id=nZeVKeeFYf9

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022),https: //openreview.net/forum?id=nZeVKeeFYf9

    2022

  12. [20]

    arXiv preprint arXiv:2108.06098 (2021)

    Hyeon-Woo, N., Ye-Bin, M., Oh, T.H.: Fedpara: Low-rank hadamard product for communication-efficient federated learning. arXiv preprint arXiv:2108.06098 (2021)

    2021 arXiv

  13. [21]

    Journal of Magnetic Res- onance Imaging 27(4), 685–691 (2008)

    Jack Jr, C.R., Bernstein, M.A., Fox, N.C., Thompson, P., Alexander, G., Harvey, D., Borowski, B., Britson, P.J., Whitwell, J.L., Ward, C., et al.: The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods. Journal of Magnetic Res- onance Imaging 27(4), 685–691 (2008)

    2008

  14. [22]

    In: AAAI

    Jie, S., Deng, Z.H.: Fact: Factor-tuning for lightweight adaptation on vision trans- former. In: AAAI. vol. 37, pp. 1060–1068 (2023)

    2023

  15. [23]

    arXiv preprint arXiv:1412.6980 (2014)

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

    2014 arXiv

  16. [24]

    Progress in Neurobiology95(4), 629–635 (2011)

    Marek, K., Jennings, D., Lasch, S., Siderowf, A., Tanner, C., Simuni, T., Coffey, C., Kieburtz, K., Flagg, E., Chowdhury, S., et al.: The Parkinson progression marker initiative (PPMI). Progress in Neurobiology95(4), 629–635 (2011)

    2011

  17. [25]

    Journal of Neurology, Neurosurgery & Psychiatry 83(2), 188–194 (2012)

    Melzer, T.R., Watts, R., MacAskill, M.R., Pitcher, T.L., Livingston, L., Keenan, R.J., Dalrymple-Alford, J.C., Anderson, T.J.: Grey matter atrophy in cognitively impaired Parkinson’s disease. Journal of Neurology, Neurosurgery & Psychiatry 83(2), 188–194 (2012)

    2012

  18. [26]

    In: ISBI

    Packhäuser, K., Folle, L., Thamm, F., Maier, A.: Generation of anonymous chest radiographs using latent diffusion models for training thoracic abnormality classi- fication systems. In: ISBI. pp. 1–5. IEEE (2023)

    2023

  19. [27]

    In: MICCAI Workshop

    Pinaya, W.H., Tudosiu, P.D., Dafflon, J., Da Costa, P.F., Fernandez, V., Nachev, P., Ourselin, S., Cardoso, M.J.: Brain imaging generation with latent diffusion models. In: MICCAI Workshop. pp. 117–126. Springer (2022)

    2022

  20. [28]

    In: MICCAI

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: MICCAI. pp. 234–241. Springer (2015)

    2015

  21. [29]

    In: ICML

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: ICML. pp. 2256–2265. PMLR (2015)

    2015

  22. [30]

    PLoS medicine 12(3), e1001779 (2015)

    Sudlow, C., Gallacher, J., Allen, N., Beral, V., Burton, P., Danesh, J., Downey, P., Elliott, P., Green, J., Landray, M., et al.: UK biobank: an open access resource for TenVOO: PEFT of 3D DDPM for MRI Image Generation Using TNs 11 identifying the causes of a wide range of com...

    2015

  23. [31]

    In: Asilomar Conference on Signals, Systems & Computers

    Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: Asilomar Conference on Signals, Systems & Computers. vol. 2, pp. 1398–1402. IEEE (2003)

    2003

  24. [32]

    In: MIDL

    Wu, J., Fu, R., Fang, H., Zhang, Y., Yang, Y., Xiong, H., Liu, H., Xu, Y.: Med- segdiff: Medical image segmentation with diffusion probabilistic model. In: MIDL. pp. 1623–1639. PMLR (2024)

    2024

  25. [33]

    In: ICLR (2024), https://openreview.net/forum?id=wfzXa8e783

    Yeh, S.Y., Hsieh, Y.G., Gao, Z., Yang, B.B., Oh, G., Gong, Y.: Navigating text- to-image customization: From lyCORIS fine-tuning to model evaluation. In: ICLR (2024), https://openreview.net/forum?id=wfzXa8e783

    2024

  26. [34]

    arXiv preprint arXiv:2106.10199 (2021)

    Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.