Pith. sign in

REVIEW 5 major objections 6 minor 53 references

This paper argues that a decoder-only autoencoder learning triplane representations, paired with a triplane-aware cross-attention diffusion model, generates high-resolution 3D medical volumes more faithfully and with less memory than encode

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 03:37 UTC pith:FNOXBLYJ

load-bearing objection Practical triplane diffusion for 3D medical volumes, but the reconstruction claim conflates per-sample fitting with learned encoding. the 5 major comments →

arxiv 2607.13812 v1 pith:FNOXBLYJ submitted 2026-07-15 eess.IV cs.CV

TCAM-Diff: Triplane-Aware Cross-Attention Medical Diffusion Model

Zhenkai Zhang , Krista A. Ehinger , Tom Drummond This is my paper
classification eess.IV cs.CV
keywords triplane representationdecoder-only autoencoderlatent diffusion3D medical image generationcross-attentionhigh-resolution medical imagingWasserstein distanceMRI/CT synthesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

TCAM-Diff tries to establish that three orthogonal 2D feature planes—a triplane representation—can compactly encode dense 3D medical volumes, and that a decoder-only autoencoder plus a cross-attention diffusion model can reconstruct and generate those volumes at resolutions (up to 512×512×512) where encoder-decoder baselines run out of memory. The paper reports better reconstruction (lower MSE, higher 3D SSIM) and better generation (lower W-GAN critic Wasserstein distance) than two baselines with similar latent dimensions. The decoder-only design replaces a learned encoder with per-volume iterative backpropagation, trading slower encoding for faster decoding and a much smaller memory footprint. If correct, this points to a practical route for generating high-resolution medical imaging data without prohibitive GPU memory.

Core claim

The central claim is that the triplane representation, until now used mainly for surfaces of 3D objects, can effectively represent dense medical volumes, and that combining a decoder-only autoencoder with a triplane-aware cross-attention diffusion model outperforms encoder-decoder latent diffusion approaches. The autoencoder stores three axis-aligned feature planes and decodes queried point features through a shared MLP; per-volume triplane features are obtained by iterative backpropagation rather than a forward encoder. The diffusion model processes each plane with 2D convolutions and uses cross-attention to exchange information across planes, and is trained to estimate both the noiseless t

What carries the argument

The central object is the triplane representation: three orthogonal 2D feature planes (Fxy, Fyz, Fxz) that together encode a 3D volume, decoded by a shared MLP. A decoder-only autoencoder learns these planes by optimizing them and the MLP against voxel intensities (using SmoothL1, perceptual, and total-variation losses), replacing the encoder with iterative backpropagation. The second stage is a diffusion model with per-plane convolutional processing and cross-attention transformer layers that let the three planes exchange information; it is trained to jointly estimate the image and the noise. The W-GAN critic Wasserstein distance is used as a distributional metric for generative quality.

Load-bearing premise

The load-bearing premise is that iterative backpropagation to fit triplane features to a given volume is a fair and reliable substitute for a learned encoder; if the reported reconstructions are produced by optimizing on the target volumes themselves, the MSE/SSIM advantage measures fitting capacity rather than generalization to unseen data.

What would settle it

Run the reconstruction procedure on a held-out volume using only the paper's described optimization (no ground-truth volume access during latent fitting) and compare with a forward-encoder variant; if the quality gap to the baselines disappears or the optimization fails to converge for many volumes, the central comparison lacks generalization. Additionally, verifying that per-volume optimization converges consistently across all test volumes would test the method's reliability.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The decoder-only autoencoder can train on a single GPU and process volumes up to 512^3, a scale at which the compared encoder-decoder latent diffusion models run out of memory.
  • Cross-attention between the three triplane planes improves generated-volume quality over concatenating planes, as shown by the paper's ablation.
  • The same triplane structure handles multi-channel (4-modality) data while preserving enough information for downstream tumour segmentation.
  • The W-GAN critic provides a distributional comparison metric for domains where established metrics like FID are impractical.
  • The triplane representation is shown to be viable for dense volumetric data, not only object surfaces.

Where Pith is reading between the lines

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

  • The reconstruction numbers are best interpreted as per-sample fitting capacity unless the paper shows how triplane latents are obtained for unseen volumes; a true encoder comparison would settle whether the gain generalizes.
  • The method suggests a memory-scaling path beyond 512^3; because decode speed is emphasized, very high-resolution generation could be obtained by increasing plane resolution and decoder capacity.
  • The cross-attention design could be adapted for conditional generation (e.g., given anatomical priors or tumour masks) by injecting conditioning into the attention blocks.
  • The Wasserstein-distance metric should be paired with structural or clinical metrics on generated volumes before drawing conclusions about generative fidelity, since a critic can be gamed by low-variance outputs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes TCAM-Diff, a two-stage generative model for high-resolution 3D medical images. In stage one, a "decoder-only autoencoder" represents each volume by three learnable 2D feature planes (triplanes) and a shared MLP decoder; the triplanes are optimized per volume by iterative backpropagation with SmoothL1, perceptual, and total-variation losses, plus RMSNorm and noise regularization. In stage two, a diffusion model operates independently on each plane and uses a cross-attention block to exchange information across planes; the training objective simultaneously estimates the clean triplane and the noise. Experiments on BraTS 128^3, Pancreas 256^3, and Colon 512^3 report reconstruction MSE/3D SSIM, downstream segmentation performance on reconstructed BraTS volumes, and generation quality measured by a W-GAN critic's Wasserstein distance. The paper claims superior reconstruction and generation compared with VAE-GAN- and VQ-GAN-based latent diffusion models at similar latent dimensions, with lower GPU memory and training time.

Significance. If the claims are substantiated, the decoder-only triplane representation is an interesting direction for memory-efficient 3D medical image generation, and the cross-plane attention mechanism is a sensible architectural contribution. The reported memory savings (1 GPU vs. 4 GPUs for stage one) and ability to handle 256^3 and 512^3 volumes are practically relevant. The proposed W-GAN critic evaluation, if made fully precise, could be a useful supplement to FID in small-sample medical settings. However, the central reconstruction claim currently conflates per-sample latent fitting with learned generalization, and the evaluation protocol for the W-GAN critic is underspecified. These issues are load-bearing for the paper's main comparison and must be addressed before the claims can be accepted.

major comments (5)
  1. [Methodology, Triplane Representation Learning; Table 1] The method defines per-object parametric triplanes F^(i) that are optimized jointly with a shared MLP decoder; there is no encoder. The paper says encoding uses iterative backpropagation, but it never specifies how a test volume is encoded: initialization, optimization steps, learning rate, convergence criterion, or validation protocol. It is also not stated whether the MSE/SSIM values in Table 1 are computed on training volumes, held-out volumes, or both. Since the triplane latent has 516,096 dimensions and is fitted per volume, the reported advantage over VAE-GAN/VQ-GAN may reflect per-sample fitting capacity rather than learned generalization. This directly undermines the claim of 'better reconstruction than encoder-decoder methods with similar-sized latent spaces.' Please specify the full test-time encoding protocol and report reconstruction on held-out volumes with the stage-one dec
  2. [Table 1 and Ablation Experiments, Table 4] The 'Parameters' column for Ours in Table 1 lists 516,096 = 3×(42×64×64), which is the per-volume latent size. The stage-one model actually stores and optimizes N such triplane sets for N training volumes, plus the shared MLP weights; these are not counted. For the baselines, the listed 524,288 is likewise the latent code size, not the total number of network parameters. Thus the comparison is at best a match of per-sample latent dimensionality, not of total trainable parameters. Please report both latent dimension and total trainable parameters (including all per-volume triplanes and the decoder) for each model, and adjust the fairness claims accordingly. Table 4's 'Params Ratio' has the same ambiguity.
  3. [Evaluation Metrics and Table 3] The W-GAN critic evaluation is not described reproducibly. The paper states that a W-GAN critic is used to estimate Wasserstein distance but does not say how the critic is trained, on which real dataset, with how many generated samples, whether the same critic is used for all models, or how the mean±std over three seeds is computed. It is also unclear whether the critic operates on decoded 3D volumes or on some other representation; for our method, generated triplanes must be decoded by the stage-one decoder before comparison. Without this information, Table 3's distances (e.g., Ours 36±4 vs. 120±20) cannot be assessed for fairness or statistical meaning. Please provide the complete evaluation protocol and, ideally, an additional established metric (or a clearly justified reason for omitting it).
  4. [Loss Functions, Eq. (4)] The perceptual loss in Eq. (4) uses a pre-trained 'squeeze' network, which is a 2D image network, but the model operates on 3D volumes. The paper does not specify how 2D features are obtained from 3D data: axial slices, orthogonal projections, or features from the triplane planes themselves. This is not merely a presentation detail, because the perceptual term is part of the reconstruction objective and can materially affect the reported MSE/SSIM. Please describe the exact feature-extraction procedure for 3D inputs and state which layers are used.
  5. [Experimental Results, Table 2] The downstream segmentation evaluation is underspecified. It is not stated whether the SegResNet segmentation model was trained on original volumes, on reconstructed volumes, or on both, nor whether the same trained segmentation model is used for all compared reconstructions. VAE-GAN is excluded due to OOM, so the comparison is only against VQ-GAN; the claim that our model 'preserves critical information for medical segmentation' should be scoped to that baseline. Please specify the training/evaluation protocol and, if feasible, include an additional baseline that can run at the same resolution.
minor comments (6)
  1. [Abstract] The phrase 'leverages generalization operations to prevent overfitting' is vague; the body names RMSNorm and noise regularization. Please state these explicitly in the abstract or the introduction.
  2. [Methodology, Eq. (9)] In the diffusion objective, the notation 'cx0' should be '\hat{x}_0' or similar, and the loss is written as a sum of two expectation terms without weighting. Please clarify how the image-estimation and noise-estimation terms are balanced and how this relates to the noise schedule.
  3. [Implementation Details] The paper says 'Additional hyperparameters, architectures, and experimental details are provided in the appendix,' but no appendix is present in the arXiv v1 text. Either include the appendix or remove the reference.
  4. [Table 4] The heading 'Params Ratio' is inconsistent with Table 1's 'Parameters' column. Since the quantity is actually latent dimension ratio, rename it to avoid confusion.
  5. [Figure 7] The diffusion ablation is presented only qualitatively. A quantitative comparison (e.g., MSE/SSIM on reconstructed volumes or the W-GAN critic distance) would make the cross-attention benefit more convincing.
  6. [Related Work / Evaluation Metrics] The Wasserstein distance formula contains typographical issues ('sup∥f∥ L≤1' and the Lipschitz constraint notation). Please correct the notation and cite a precise source for the practical estimator used.

Circularity Check

1 steps flagged

Stage-one reconstruction MSE/SSIM is computed on volumes whose triplane latents were optimized against those same volumes, so the 'better reconstruction' claim reduces to per-sample fitting capacity.

specific steps
  1. fitted input called prediction [Methodology – Triplane Representation Learning; Loss Functions and Regularization Approaches; Table 1]
    "We predefined multiple parametric triplane representations, {Fxy(0),Fyz(0),Fxz(0)},..., {Fxy(i),Fyz(i),Fxz(i)},... based on the number of objects in the training set. ... During the training process of the triplane decoder-only model, we optimize both the MLP network and the parameters of the triplane feature planes. ... Table 1: Quantitative comparison of reconstruction quality ... Ours 0.0008 0.9311"

    The decoder-only 'autoencoder' has no encoder: the per-object triplanes F^(i) are free parameters optimized (Eq. 8) against the very volume whose reconstruction is scored in Table 1. Reported MSE is the terminal value of the fitting objective for each sample, not an out-of-sample prediction by a learned encoder. The paper does not state a train/test split or test-time encoding procedure; if a test volume is used, its triplane is still fitted to that same volume by iterative backpropagation. Comparing this per-sample fitting capacity with VAE-GAN/VQ-GAN, whose amortized encoders are evaluated without per-sample optimization, makes the claimed reconstruction superiority partly circular by construction.

full rationale

The diffusion stage (Eq. 9, cross-attention, W-GAN critic evaluation) is not circular: the W-GAN critic is an external trained metric, and the generative comparison does not reduce to the fitting loss. The only significant circular step is the stage-one autoencoder evaluation. Because the central abstract claim includes 'better reconstruction... than other encoder-decoder methods', and that reconstruction score is the objective minimized per sample, the paper's headline reconstruction result is partly forced by construction rather than by learned generalization. No load-bearing self-citation or imported uniqueness theorem was found. Score 6 reflects one central 'prediction' (reconstruction quality) reducing to a per-sample fit, while the generation claims retain independent content.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The method is a composition of known components (triplanes, autodecoding, latent diffusion, cross-attention) with several hand-tuned hyperparameters. Its main empirical claims rest on per-sample fitted latents and a fitted critic. No new physical entities are introduced; the only new objects are model components.

free parameters (7)
  • SmoothL1 beta = 0.3
    Set by hand in Eq. 3; controls the transition between L1 and L2 behavior and affects reconstruction sharpness.
  • MLP noise scale C = 0.32
    Eq. 7; authors state C=0.32 'can achieve better generalization in our experiments', i.e. tuned on validation rather than derived.
  • Perceptual loss weight lambda_1 = 1e-2 (BraTS); 0 (higher-res)
    Eq. 8; chosen by hand and disabled for high-resolution datasets, changing the training objective across experiments.
  • TV regularization weight lambda_2 = 1e-3
    Eq. 8; chosen by hand.
  • Triplane channel count = 42 in main model; ablation from 18 to 48
    Table 1/4; architecture choice controlling latent size, tuned to approximately match baseline latent dimensions.
  • Learning rate = 3e-5 (BraTS), 1e-5 (Pancreas/Colon)
    Implementation details; per-dataset hand-set hyperparameter.
  • Gradient clipping max norm = 1.0
    Implementation details; stabilizing choice not derived from the problem.
axioms (5)
  • domain assumption A point's intensity in a dense 3D volume is determined by concatenated features sampled from its three axis-aligned plane projections and decoded by a shared MLP.
    Triplane Representation Learning section; this representational assumption is central to the method and is not proven for dense medical volumes.
  • domain assumption All volumes in a dataset share a single MLP decoder; intensity mapping is consistent across patients.
    Triplane Representation Learning section; needed for the shared decoder and for generalization across the dataset.
  • domain assumption Per-object triplane features optimized via backpropagation converge to a representation that supports diffusion training.
    Decoder-only training objective Eq. 2/8; no convergence guarantee or analysis is provided.
  • domain assumption The W-GAN critic provides a reliable estimate of Wasserstein distance for comparing generative models.
    Evaluation Metrics section; a trained critic gives only a lower bound unless fully optimized under the Lipschitz constraint, and no cross-model calibration is shown.
  • domain assumption Standard diffusion denoising assumptions apply to RMSNorm-normalized triplane latents (e.g., near-Gaussian marginal).
    Diffusion model section; no distributional analysis of the learned latents is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 11414 in / 15735 out tokens · 144056 ms · 2026-08-02T03:37:02.955116+00:00 · methodology

0 comments
read the original abstract

We introduce TCAM-Diff, a novel 3D medical image generation model that reduces the memory requirements to encode and generate high-resolution 3D data. This model utilizes a decoder-only autoencoder method to learn triplane representation from dense volume and leverages generalization operations to prevent overfitting. Subsequently, it uses a triplane-aware cross-attention diffusion model to learn and integrate these features effectively. Furthermore, the features generated by the diffusion model can be rapidly transformed into 3D volumes using a pre-trained decoder module. Our experiments on three different scales of medical datasets, BrainTumour 128 x 128 x 128, Pancreas 256 x 256 x 256, and Colon 512 x 512 x 512, demonstrate outstanding results. We utilized MSE and SSIM to assess reconstruction quality and leveraged the Wasserstein Generative Adversarial Network (W-GAN) critic to assess generative quality. Comparisons with existing approaches show that our method gives better reconstruction and generation results than other encoder-decoder methods with similar-sized latent spaces.

Figures

Figures reproduced from arXiv: 2607.13812 by Krista A. Ehinger, Tom Drummond, Zhenkai Zhang.

Figure 1
Figure 1. Figure 1: The overview of our two-stage model architecture. (a). Triplane representation learning: triplane features [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The attention block for Fxy, incorporating both self-attention and cross-attention mechanisms. The same structure is applied to the feature planes Fyz and Fxz. The structure ensures the effective integration of information across different planes. 2023) propose using the total variation (TV) regularization terms can simplify the data manifold by eliminating un￾necessary high-frequency information from feat… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of reconstruction qual [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Downstream segmentation task: evaluation of re [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance between the diffusion model with [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

53 extracted references · 1 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2017 , eprint=

    Attention Is All You Need , author=. 2017 , eprint=

    2017

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    Communications of the ACM , volume=

    Generative adversarial networks , author=. Communications of the ACM , volume=. 2020 , publisher=

    2020

  13. [13]

    Foundations and Trends

    An introduction to variational autoencoders , author=. Foundations and Trends. 2019 , publisher=

    2019

  14. [14]

    Journal of Machine Learning Research , volume=

    Normalizing flows for probabilistic modeling and inference , author=. Journal of Machine Learning Research , volume=

  15. [15]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  16. [16]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  17. [17]

    IEEE Transactions on Medical Robotics and Bionics , volume=

    Data augmentation of 3D brain environment using deep convolutional refined auto-encoding alpha GAN , author=. IEEE Transactions on Medical Robotics and Bionics , volume=. 2020 , publisher=

    2020

  18. [18]

    arXiv preprint arXiv:2307.11926 , year=

    PartDiff: image super-resolution with partial diffusion models , author=. arXiv preprint arXiv:2307.11926 , year=

    Pith/arXiv arXiv

  19. [19]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    3d neural field generation using triplane diffusion , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  20. [20]

    Advances in Neural Information Processing Systems , volume=

    Autodecoding latent 3d diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  21. [21]

    arXiv preprint arXiv:2309.15103 , year=

    Lavie: High-quality video generation with cascaded latent diffusion models , author=. arXiv preprint arXiv:2309.15103 , year=

    Pith/arXiv arXiv

  22. [22]

    ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Diffvoice: Text-to-speech with latent diffusion , author=. ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2023 , organization=

    2023

  23. [23]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    Generation of 3D brain MRI using auto-encoding generative adversarial networks , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2019 , organization=

    2019

  24. [24]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Efficient geometry-aware 3d generative adversarial networks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  25. [25]

    IEEE Access , volume=

    Synthesis of 3D MRI brain images with shape and texture generative adversarial deep neural networks , author=. IEEE Access , volume=. 2021 , publisher=

    2021

  26. [26]

    ArXiv , year=

    Brain Imaging Generation with Latent Diffusion Models , author=. ArXiv , year=

  27. [27]

    arXiv preprint arXiv:2211.03364 , year=

    Medical Diffusion: Denoising Diffusion Probabilistic Models for 3D Medical Image Generation , author=. arXiv preprint arXiv:2211.03364 , year=

    Pith/arXiv arXiv

  28. [28]

    ArXiv , year=

    Wasserstein GAN , author=. ArXiv , year=

  29. [29]

    ACM Computing Surveys , volume=

    Diffusion models: A comprehensive survey of methods and applications , author=. ACM Computing Surveys , volume=. 2023 , publisher=

    2023

  30. [30]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Wavelet diffusion models are fast and scalable image generators , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  31. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Class-balancing diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  32. [32]

    arXiv preprint arXiv , volume=

    Diffusion models for medical image analysis: a comprehensive survey , author=. arXiv preprint arXiv , volume=

  33. [33]

    3D-StyleGAN: A style-based generative adversarial network for generative modeling of three-dimensional medical images , author=. Deep Generative Models, and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings ...

    2021

  34. [34]

    arXiv preprint arXiv:2101.05218 , year=

    Three dimensional mr image synthesis with progressive generative adversarial networks , author=. arXiv preprint arXiv:2101.05218 , year=

    Pith/arXiv arXiv

  35. [35]

    Vox2Vox: 3D-GAN for brain tumour segmentation , author=. Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 6th International Workshop, BrainLes 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Revised Selected Papers, Part I 6 , pages=. 2021 , organization=

    2020

  36. [36]

    Medical Imaging with Deep Learning , year=

    Three-dimensional medical image synthesis with denoising diffusion probabilistic models , author=. Medical Imaging with Deep Learning , year=

  37. [37]

    Proceedings of the IEEE international conference on computer vision , pages=

    Fast r-cnn , author=. Proceedings of the IEEE international conference on computer vision , pages=

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    Root mean square layer normalization , author=. Advances in Neural Information Processing Systems , volume=

  39. [39]

    Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14 , pages=

    Perceptual losses for real-time style transfer and super-resolution , author=. Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14 , pages=. 2016 , organization=

    2016

  40. [40]

    Asian Conference on Machine Learning , year=

    Improving Denoising Diffusion Models via Simultaneous Estimation of Image and Noise , author=. Asian Conference on Machine Learning , year=

  41. [41]

    IEEE transactions on medical imaging , volume=

    The multimodal brain tumor image segmentation benchmark (BRATS) , author=. IEEE transactions on medical imaging , volume=. 2014 , publisher=

    2014

  42. [42]

    arXiv preprint arXiv:1811.02629 , year=

    Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the BRATS challenge , author=. arXiv preprint arXiv:1811.02629 , year=

    Pith/arXiv arXiv

  43. [43]

    B. H. Menze and A. Jakab and S. Bauer and J. Kalpathy-Cramer and K. Farahani and J. Kirby and others , title =. IEEE Transactions on Medical Imaging , year =

  44. [44]

    ArXiv , year=

    A large annotated medical image dataset for the development and evaluation of segmentation algorithms , author=. ArXiv , year=

  45. [45]

    2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    High-Resolution Image Synthesis with Latent Diffusion Models , author=. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    2022

  46. [46]

    doi:10.5281/zenodo.4323059 , url =

    Project MONAI , year =. doi:10.5281/zenodo.4323059 , url =

    work page doi:10.5281/zenodo.4323059

  47. [47]

    doi:10.48550/ARXIV.2211.03364 , url =

    Khader, Firas and Mueller-Franzes, Gustav and Arasteh, Soroosh Tayebi and Han, Tianyu and Haarburger, Christoph and Schulze-Hagen, Maximilian and Schad, Philipp and Engelhardt, Sandy and Baessler, Bettina and Foersch, Sebastian and Stegmaier, Johannes and Kuhl, Christiane and Nebelung, Sven and Kather, Jakob Nikolas and Truhn, Daniel , title =. doi:10.485...

    Pith/arXiv arXiv doi:10.48550/arxiv.2211.03364

  48. [48]

    Computer Methods and Programs in Biomedicine , volume=

    Inflating 2D convolution weights for efficient generation of 3D medical images , author=. Computer Methods and Programs in Biomedicine , volume=. 2023 , publisher=

    2023

  49. [49]

    arXiv preprint arXiv:2402.19043 , year=

    WDM: 3D Wavelet Diffusion Models for High-Resolution Medical Image Synthesis , author=. arXiv preprint arXiv:2402.19043 , year=

    Pith/arXiv arXiv

  50. [50]

    IEEE Transactions on Image Processing , year=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE Transactions on Image Processing , year=

  51. [51]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  52. [52]

    arXiv preprint arXiv:1904.00625 , year=

    Med3d: Transfer learning for 3d medical image analysis , author=. arXiv preprint arXiv:1904.00625 , year=

    Pith/arXiv arXiv 1904

  53. [53]

    3D MRI brain tumor segmentation using autoencoder regularization , author=. Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 4th International Workshop, BrainLes 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16, 2018, Revised Selected Papers, Part II 4 , pages=. 2019 , organization=

    2018