Pith. sign in

REVIEW 2 major objections 6 minor 27 references

Forcing void tokens to attend only known-healthy tissue, with a contralateral bias and mirrored-hemisphere input, drives a volumetric U-DiT to BraTS-2026 healthy-region SSIM 0.864.

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 · grok-4.5

2026-07-31 21:39 UTC pith:5BW66HHV

load-bearing objection Solid BraTS methods paper: healthy-only bottleneck attention is the real lever, ablations and official scores back it, novelty is moderate but usable. the 2 major comments →

arxiv 2607.27974 v1 pith:5BW66HHV submitted 2026-07-30 cs.CV

Now You Have My Healthy Attention: A U-DiT for Brain-MRI Inpainting

Danilo Danese , Angela Lombardi , Tommaso Di Noia This is my paper
classification cs.CV
keywords InpaintingBraTS 2026MRITransformersU-DiThealthy-only attentioncontralateral symmetrybrain MRI
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.

The paper sets out to show that brain-MRI healthy-tissue inpainting, scored by distortion metrics, is improved less by a generic transformer backbone than by inductive biases that match the task. The authors build a deterministic volumetric U-DiT: convolutions and skip connections keep local detail, while one downsampled global self-attention block with 3D rotary embeddings supplies long-range context. Two biases carry the gains. Occluded tokens are barred from attending other occluded tokens and may read only known-healthy tokens of the same volume, with a learned preference for the contralateral homologue. Separately, the network is given the patient’s mirrored healthy hemisphere plus a validity map as extra input channels. On the official BraTS-2026 validation set of 219 cases the submission reaches mean healthy-region SSIM 0.864, PSNR 24.7 dB and MSE 4.6×10⁻³. A reader cares because clinical tumor scans almost never include a patient-specific healthy reference, and the same observed-only and symmetry priors are reusable wherever holes must be filled from visible anatomy.

Core claim

Two task-specific inductive biases—healthy-only bottleneck attention (void tokens attend exclusively to known-healthy tokens, with a learned bias toward each query’s contralateral homologue) and a contralateral-symmetry input that supplies the mirrored healthy hemisphere as a patient-specific prior—are the main drivers that raise a volumetric U-DiT regression model to official BraTS-2026 validation scores of mean healthy-region SSIM 0.864, PSNR 24.7 dB and MSE 4.6×10⁻³ over 219 cases, outperforming unconstrained and full-transformer alternatives on the distortion metrics the challenge uses.

What carries the argument

Healthy-only attention with contralateral weighting: a parameter-free −∞ key bias that drops predominantly void tokens from the attendable set, plus a learned per-head Gaussian preference peaked at the query’s mid-sagittal mirror, used together with a four-channel contralateral-symmetry input (voided volume, mask, mirror intensities, validity).

Load-bearing premise

The method assumes the brain is roughly left–right symmetric and lesions usually sit on one side, so the estimated mid-sagittal mirror is a reliable patient-specific template for the missing tissue.

What would settle it

On the official validation distribution, remove the healthy-only key bias and the contralateral input/attention term from the submitted U-DiT; if healthy-region SSIM, PSNR and MSE stay essentially unchanged versus the plain backbone, the claim that these biases drive the leaderboard scores is false.

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

If this is right

  • Any coarse-attention inpainting network can force occluded queries to read only observed tissue via a parameter-free key bias, without redesigning the backbone.
  • A patient-specific contralateral template still improves distortion metrics even when mirror test-time augmentation is already applied.
  • Distortion-optimal regression will keep suppressing high-frequency texture; residual generative refinement is the route the authors identify to regain anatomical realism without losing SSIM/PSNR/MSE.
  • The same healthy-only and bilateral-symmetry priors should transfer to other unilateral-lesion medical inpainting tasks scored on distortion metrics.

Where Pith is reading between the lines

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

  • Stratifying leaderboard error by midline-crossing and bilateral lesions would quantify how often the symmetry prior fails even with the validity gate.
  • Healthy-only key masking is architecture-agnostic and could be inserted into diffusion or flow-matching inpainters to reduce hallucination from other missing regions while still sampling texture.
  • Leaderboards that report only SSIM/PSNR/MSE will systematically prefer smooth conditional-mean regressors unless a texture- or perception-aware secondary score is added.

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

2 major / 6 minor

Summary. The paper proposes a volumetric U-DiT regression network for the BraTS Local Synthesis (healthy-tissue inpainting) task. A convolutional encoder–decoder carries local detail; a single bottleneck self-attention block on a 2× space-to-channel downsampled grid (3D RoPE, cosine-similarity attention) supplies long-range context. Two task-specific inductive biases are introduced: (i) a parameter-free healthy-only key mask that forces void queries to attend only to known-healthy tokens, plus a learned contralateral logit bias (Eq. 2); (ii) a contralateral-symmetry input (mirrored voided volume and validity channel; Eqs. 3–4). Training uses masked L1+SSIM plus a light high-pass term, random-mask augmentation, EMA, and mirror TTA at inference. Cumulative ablations on an internal split and official BraTS-2026 validation scores over 219 cases (SSIM 0.864, PSNR 24.7 dB, MSE 4.6×10⁻³) are reported, together with an explicit discussion of residual smoothness under distortion-optimal regression.

Significance. If the reported gains hold, the work supplies two reusable, low-cost inductive biases for medical inpainting—especially the parameter-free healthy-only attention constraint, which the ablation identifies as the single largest lift and which applies to any coarse-attention inpainting backbone. The architecture choice (conv U-Net + one downsampled attention block rather than a full 3D transformer) is well motivated by the low-frequency nature of the required non-local context and by a clear negative result for a full-transformer baseline (~0.58 SSIM). Code is released, official-server means and standard deviations are given, and the authors candidly analyse the perception–distortion trade-off rather than overselling texture realism. The contribution is incremental relative to prior BraTS inpainting winners and symmetry priors, but the combination is cleanly engineered and the empirical evidence is transparent.

major comments (2)
  1. [§3.2, Table 1] Table 1 (top block): the ablation is strictly cumulative and ends with a parenthetical row “+ TTA & annealing (no contralateral attention)” that is hard to interpret. The text claims healthy-only attention is “the most effective single change” and that contralateral input yields a further distortion-metric gain, yet there is no factorial (or leave-one-component-out) measurement that isolates (a) healthy-only key masking alone, (b) the learned contralateral logit bias λ_h/σ in Eq. (2) alone, and (c) the contralateral input channels alone, each with and without TTA. Without those cells, the load-bearing attribution in §3.2 and the abstract remains only partially supported. A small revised table or appendix with independent ablations on the same internal split would make the central claim fully checkable.
  2. [§3.1–3.2] §2.1 and §3.1–3.2 never state the size or construction of the internal held-out split on which the 0.848→0.865 SSIM ladder is measured. Gains of 0.003–0.005 SSIM and ~0.0002–0.0004 MSE are small relative to the official-set standard deviations (SSIM std 0.085, MSE std 0.0032). Reporting N, the split protocol, and ideally bootstrap or paired tests would let readers judge whether the ranked order of components is stable. This does not overturn the official-server result, but it is required for the ablation-driven design narrative.
minor comments (6)
  1. [§2.3, Eq. (2)] Eq. (2): the symbol ¯v_j is used for the pooled mirror-validity gate, while §2.4 uses ¯v_h for the per-voxel validity map. A single consistent notation (and an explicit statement that the gate is average-pooled onto the token grid) would avoid confusion.
  2. [Fig. 1] Fig. 1 caption and the “× 2 Residual blocks” annotation are slightly ambiguous about how many U-DiT blocks sit at the bottleneck (text says “a single” block). Clarify.
  3. [§2.5] §2.5: the high-pass loss L_hf and its weight λ_hf are mentioned but never defined. A one-line formula (e.g., masked L1 on a Laplacian or high-frequency residual) belongs in the main text or appendix.
  4. [Abstract / §5 / Table 1] Official leaderboard MSE is written 4.6×10^{-3} in the abstract and 4.57×10^{-3} in the conclusion; Table 1 reports 0.0046. Harmonise to a single rounding.
  5. [§1, §3.3] Related-work placement of the 2024/2025 BraTS inpainting winners [26,27] is appropriate, but a short quantitative comparison (their published validation numbers vs. yours, with the caveat of possibly different validation sets) would help situate the 0.864 SSIM result.
  6. [Abstract, §2.4] Typos / style: “controlateral” → “contralateral” (p. 3); “TheASNR-MICCAI” spacing in the abstract opening; “whP” line-break artefact in Eq. (3).

Circularity Check

0 steps flagged

No significant circularity: empirical BraTS methods result evaluated on an external leaderboard, with ablations that do not reduce to fitted inputs by construction.

full rationale

The paper’s central claim is an empirical architecture-and-inductive-bias result for BraTS-2026 Task-4 inpainting: a convolutional U-DiT backbone plus parameter-free healthy-only key masking (Eq. 2) and contralateral-symmetry input channels (Eqs. 3–4) reaches official validation SSIM/PSNR/MSE on 219 cases. Training optimizes masked MAE+SSIM on random voids; evaluation is on a held-out internal split and the official challenge server. The healthy-only bias is a fixed −∞ mask on void keys, not a quantity fitted to the reported score. Contralateral λ_h/σ are learned network parameters, not post-hoc fits renamed as predictions. Self-citation to the authors’ WaveDiT [8] appears only as a cheaper underperforming baseline in Table 1, not as a load-bearing uniqueness or uniqueness-of-form premise. Bilateral-symmetry priors are cited to external work [15]. Nothing in the derivation chain equates a claimed prediction to its own input by construction. Score 0 is the appropriate honest finding.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The work is an empirical deep-learning methods paper. Load-bearing content is almost entirely domain assumptions about brain anatomy and the BraTS scoring protocol, plus ordinary architectural hyperparameters. No new physical entities are postulated. The central performance claim rests on the symmetry prior, the healthy-only attention rule, and standard supervised regression on challenge data.

free parameters (5)
  • per-head contralateral bias strength λ_h and width σ = learned (not numerically reported)
    Learned parameters in the attention logit (Eq. 2) that control how strongly each head prefers the mirror homologue; zero-init friendly but fitted during training.
  • high-pass loss weight λ_hf = not reported
    Scalar balancing the optional high-frequency term against masked MAE and SSIM; chosen to discourage over-smoothing.
  • learnable per-head attention temperature s = learned (not numerically reported)
    Scales cosine-similarity logits inside the bottleneck attention; fitted end-to-end.
  • void-fraction key-mask threshold 0.5 = 0.5
    Hard rule that any coarse token with pooled void fraction >0.5 receives −∞ key bias; hand-chosen cutoff.
  • architecture and training hyperparameters (base width 64, r=2 downsample, crop 144×208×208, AdamW + cosine schedule, EMA = as stated in §2.2 and §2.5
    Standard but claim-relevant knobs that determine capacity and the reported scores; not derived from theory.
axioms (5)
  • domain assumption BraTS scoring on healthy sub-region of the mask by SSIM, PSNR, MSE is the objective; conditional-mean regression is therefore the appropriate model class.
    Stated in abstract and §1; drives the choice of deterministic regression over generative sampling.
  • domain assumption Healthy human brains are approximately bilaterally symmetric and lesions in this cohort are typically unilateral, so the contralateral hemisphere is a useful patient-specific template.
    Invoked in §1, §2.3–2.4 and citations to [15]; underpins both the mirror input and the contralateral attention bias.
  • ad hoc to paper Non-local information needed for inpainting is predominantly low-frequency and can be captured by a single global-attention block at a 2× downsampled bottleneck.
    Design rationale in §2.2; justified by preliminary full-transformer failure (0.58 SSIM) but not proved in general.
  • ad hoc to paper Mid-sagittal plane can be adequately estimated as the intensity-weighted centroid along the left–right axis of the voided volume (Eq. 3).
    §2.4; simpler than a dedicated MSP detector and shared by both the input channel and the attention bias.
  • standard math Standard facts of multi-head attention, RoPE, U-Net skip connections, AdamW, EMA, and SSIM as a differentiable loss.
    Background tools cited and used without modification.
invented entities (2)
  • Healthy-only attention with contralateral logit bias (Eq. 2) independent evidence
    purpose: Force void queries to read only known-healthy keys and prefer the anatomical homologue, removing fixed-point coupling among missing tokens.
    Architectural mechanism introduced by the paper; not a physical entity, but the main postulated inductive bias. independent_evidence is true in the weak sense that the ablation isolates its metric contribution on held-out data.
  • Validity-gated contralateral-symmetry input channels independent evidence
    purpose: Supply a patient-specific mirror template and a mask of where that template is defined.
    Input featurization specific to this work; ablation shows a smaller distortion-metric gain. Falsifiable via the same held-out metrics.

pith-pipeline@v1.2.0-daily-grok45 · 14758 in / 3816 out tokens · 77537 ms · 2026-07-31T21:39:17.158369+00:00 · methodology

0 comments
read the original abstract

The ASNR-MICCAI BraTS Local Synthesis (Inpainting) task asks for the anatomically plausible completion of healthy brain tissue within a masked region of a T1-weighted MRI, providing a tumor-free anatomical reference for downstream analysis. As the task is scored by distortion metrics (SSIM, PSNR, MSE), we build a deterministic regression model and focus on giving it inductive biases tailored to inpainting. Our network follows the U-DiT principle of performing self-attention on a downsampled token grid: a volumetric encoder-decoder imports long-range context through a downsampled global self-attention block with three-dimensional rotary position embeddings, while convolutions and skip connections preserve high-frequency detail. Two ideas drive our results. First, we constrain the attention so that occluded ("void") tokens attend only to known-healthy tokens of the same volume, with a learned bias toward each query's contralateral homologue, forcing the completion to be inferred from observed anatomy rather than from other unknown regions. Second, we add a contralateral-symmetry input that supplies the mirrored healthy hemisphere as a patient-specific prior; since the brain is approximately bilaterally symmetric and lesions are typically unilateral, this prior improves the distortion metrics at matched structural similarity. On the official BraTS-2026 validation leaderboard our submission reaches a mean healthy-region SSIM of $0.864$, PSNR of $24.7$\,dB and MSE of $4.6{\times}10^{-3}$ over $219$ cases. We further analyse the residual smoothness inherent to distortion-optimal regression and discuss its implications for anatomical realism.

Figures

Figures reproduced from arXiv: 2607.27974 by Angela Lombardi, Danilo Danese, Tommaso Di Noia.

Figure 1
Figure 1. Figure 1: The U-DiT-style volumetric network. A convolutional encoder–decoder [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Healthy-only attention with contralateral weighting, shown for four vali [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Contralateral-symmetry input, illustrated on a double-lesion case with [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative completion for a validation case (our submitted mirror-TTA [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Perception-distortion trade-off on one case, three views. Columns, left [PITH_FULL_IMAGE:figures/full_fig_p010_5.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

27 extracted references · 5 linked inside Pith

  1. [1]

    etal.,A.K.:Federatedbenchmarkingofmedicalartificialintelligencewithmedperf. Nat. Mac. Intell.5(7), 799–810 (2023)

    2023

  2. [2]

    et al., F.K.: The brain tumor segmentation (brats) challenge: Local synthesis of healthy brain tissue via inpainting (2024), https://arxiv.org/abs/2305.08992

    Pith/arXiv arXiv 2024

  3. [3]

    et al., U.B.: The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmen- tation and radiogenomic classification (2021), https://arxiv.org/abs/2107.02314

    Pith/arXiv arXiv 2021

  4. [4]

    The Cancer Imaging Archive (2017)

    Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann, J., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for the pre-operative scans of the TCGA-GBM collection. The Cancer Imaging Archive (2017). https://doi.org/10.7937/K9/TCIA.2017.KLXWJJ1Q

    doi:10.7937/k9/tcia.2017.klxwjj1q 2017

  5. [5]

    The Cancer Imaging Archive (2017)

    Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann, J., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for the pre-operative scans of the TCGA-LGG collection. The Cancer Imaging Archive (2017). https://doi.org/10.7937/K9/TCIA.2017.GJQ7R0EF

    doi:10.7937/k9/tcia.2017.gjq7r0ef 2017

  6. [6]

    Scientific Data 4(1), 170117 (Sep 2017)

    Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features. Scientific Data 4(1), 170117 (Sep 2017). https://doi.org/10.1038/sdata.2017.117, https://doi.org/ 10.1038/sdata.2017.117

    doi:10.1038/sdata.2017.117 2017

  7. [7]

    In: ICML

    Crowson, K., Baumann, S.A., Birch, A., Abraham, T.M., Kaplan, D.Z., Shippole, E.: Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers. In: ICML. Proceedings of Machine Learning Research, vol. 235, pp. 9550–9575. PMLR / OpenReview.net (2024)

    2024

  8. [8]

    CoRRabs/2606.08670(2026)

    Danese, D., Lombardi, A., Fasano, G., Attimonelli, M., Noia, T.D.: Wavedit: Distribution-aware wavelet flow matching for efficient 3d brain MRI synthesis. CoRRabs/2606.08670(2026)

    Pith/arXiv arXiv 2026

  9. [9]

    In: DGM4MICCAI@MICCAI

    Durrer, A., Bieder, F., Friedrich, P., Menze, B.H., Cattin, P.C., Kofler, F.: fastwdm3d: Fast and accurate 3d healthy tissue inpainting. In: DGM4MICCAI@MICCAI. Lecture Notes in Computer Science, vol. 16128, pp. 171–181. Springer (2025)

    2025

  10. [10]

    In: DGM4MICCAI@MICCAI

    Durrer, A., Wolleb, J., Bieder, F., Friedrich, P., Melie-García, L., Ocampo-Pineda, M., Bercea, C.I., Hamamci, I.E., Wiestler, B., Piraud, M., Yaldizli, Ö., Granziera, C., Menze, B.H., Cattin, P.C., Kofler, F.: Denoising diffusion models for 3d healthy brain tissue inpainting. In: DGM4MICCAI@MICCAI. Lecture Notes in Computer Science, vol. 15224, pp. 87–97...

    2024

  11. [11]

    CoRR abs/2411.04630(2024)

    Ferreira, A., Luijten, G., Puladi, B., Kleesiek, J., Alves, V., Egger, J.: Brain tumour removing and missing modality generation using 3d WDM. CoRR abs/2411.04630(2024)

    Pith/arXiv arXiv 2024

  12. [12]

    In: CVPR

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 15979–15988. IEEE (2022)

    2022

  13. [13]

    In: ICLR (Poster)

    Loshchilov, I., Hutter, F.: SGDR: stochastic gradient descent with warm restarts. In: ICLR (Poster). OpenReview.net (2017) 12 D. Danese et al

    2017

  14. [14]

    In: ICLR (Poster)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (Poster). OpenReview.net (2019)

    2019

  15. [15]

    In: MICCAI (12)

    Ma, Y., Wang, D., Liu, P., Masters, L., Barnett, M., Cai, T.W., Wang, C.: Sym- metry awareness encoded deep learning framework for brain imaging analysis. In: MICCAI (12). Lecture Notes in Computer Science, vol. 15012, pp. 742–752. Springer (2024)

    2024

  16. [16]

    IEEE Transactions on Medical Imaging99(12 2014)

    Menze, B., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahaniy, K., Kirby, J., Burren, Y., Porz, N., Slotboomy, J., Wiest, R., Lancziy, L., Gerstnery, E., Webery, M.A., Arbel, T., Avants, B., Ayache, N., Buendia, P., Collins, L., Cordier, N., Van Leemput, K.: The multimodal brain tumor image segmentation benchmark (brats). IEEE Transactions on Medical Im...

    arXiv 2014

  17. [17]

    In: ICCV

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV. pp. 4172–4182. IEEE (2023)

    2023

  18. [18]

    SIAM Journal on Control and Optimization30(4), 838–855 (1992)

    Polyak, B.T., Juditsky, A.B.: Acceleration of stochastic approximation by av- eraging. SIAM Journal on Control and Optimization30(4), 838–855 (1992). https://doi.org/10.1137/0330046, https://doi.org/10.1137/0330046

    doi:10.1137/0330046 1992

  19. [19]

    In: CVPR

    Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR. pp. 1874–1883. IEEE Computer Society (2016)

    2016

  20. [20]

    Neurocomputing568, 127063 (2024)

    Su, J., Ahmed, M.H.M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced transformer with rotary position embedding. Neurocomputing568, 127063 (2024)

    2024

  21. [21]

    In: NeurIPS (2024)

    Tian, Y., Tu, Z., Chen, H., Hu, J., Xu, C., Wang, Y.: U-dits: Downsample tokens in u-shaped diffusion transformers. In: NeurIPS (2024)

    2024

  22. [22]

    Neurocomputing338, 34–45 (2019)

    Wang, G., Li, W., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T.: Aleatoric uncertainty estimation with test-time augmentation for medical image segmenta- tion with convolutional neural networks. Neurocomputing338, 34–45 (2019)

    2019

  23. [23]

    IEEE Trans

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process.13(4), 600–612 (2004)

    2004

  24. [24]

    In: ICML

    Wortsman, M., Ilharco, G., Gadre, S.Y., Roelofs, R., Lopes, R.G., Morcos, A.S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., Schmidt, L.: Model soups: averagingweightsofmultiplefine-tunedmodelsimprovesaccuracywithoutincreas- ing inference time. In: ICML. Proceedings of Machine Learning Research, vol. 162, pp. 23965–23998. PMLR (2022)

    2022

  25. [25]

    In: BraTS/CrossMoDA@MICCAI

    Zhang, J., Chen, K., Weng, Y.: Synthesis of healthy tissue within tumor area via u-net. In: BraTS/CrossMoDA@MICCAI. Lecture Notes in Computer Science, vol. 14669, pp. 233–240. Springer (2023)

    2023

  26. [26]

    In: BraTS-Lighthouse/AIMS-TBI@MICCAI (2)

    Zhang,J.,Weng,Y.,Chen,K.:Robust3dbrainMRIinpaintingwithrandommask- ing augmentation. In: BraTS-Lighthouse/AIMS-TBI@MICCAI (2). Lecture Notes in Computer Science, vol. 16377, pp. 102–109. Springer (2025), winner, BraTS- Inpainting 2025

    2025

  27. [27]

    CoRRabs/2507.18126(2025), winner, BraTS Local-Synthesis/Inpainting 2024

    Zhang, J., Weng, Y., Chen, K.: U-net based healthy 3d brain tissue inpainting. CoRRabs/2507.18126(2025), winner, BraTS Local-Synthesis/Inpainting 2024

    Pith/arXiv arXiv 2025