U-Net Based Healthy 3D Brain Tissue Inpainting

Juexin Zhang; Ke Chen; Ying Weng

REVIEW 3 major objections 5 minor 1 cited by

U-Net Based Healthy 3D Brain Tissue Inpainting

T0 review · 3 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 A standard 3D U-Net with random mask augmentation and a combined MAE and SSIM loss reconstructs healthy brain tissue at 0.841 SSIM, 23.257 PSNR, and 0.007 MSE on the BraTS-Local-Inpainting validation set, taking first place in the…

desk verdict Standard U-Net for BraTS inpainting with independent external metrics but no baselines and an unverifiable first-place claim; useful as a challenge report, not as a standalone contribution. read the letter →

arxiv 2507.18126 v1 pith:SPZGBIZ6 submitted 2025-07-24 eess.IV cs.AIcs.CV

Juexin Zhang , Ying Weng , Ke Chen This is my paper

classification eess.IVcs.AIcs.CV

keywords HealthytissuesynthesisBraTS2024U-NetInpaintingBrainMRI3DconvolutionDataaugmentationStructuralsimilarity

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper enters the BraTS-Local-Inpainting challenge and argues that a conventional 3D U-Net, given randomly generated healthy-mask variations and a loss combining mean absolute error on the inpainted region with structural similarity on the whole volume, is sufficient to synthesize healthy brain tissue where a tumor has been removed. The model is trained on T1 MRI scans with expert-approved tumor annotations and algorithm-generated healthy masks, and it reaches a validation SSIM of 0.841, a PSNR of 23.257, and an MSE of 0.007, placing first in the challenge. The result matters because it shows that a simple and easily reproduced architecture can produce synthetic healthy scans for downstream AI models, which may ease data scarcity and privacy constraints in brain-tumor analysis.

What carries the argument

The load-bearing object is a 3D U-Net with three stages of downsampling and upsampling, each block containing two 3D convolutions with kernel size 3, PReLU activations, instance normalization, and dropout in the bridge and upsampling blocks; skip connections carry features between corresponding levels. The method also uses a data-augmentation scheme that generates five healthy masks per scan and randomly mirrors and rotates them to increase the variety of regions the network must inpaint, and a loss equal to $L = \lambda_1 \mathrm{MAE}(I, GT) + \lambda_2 \mathrm{SSIM}(I, GT)$, where the MAE is taken only over voxels in the healthy mask and SSIM over the entire image. This combination carries the argument: the masks teach the network to fill arbitrary healthy regions, while the two losses balance pixel-level fidelity with structural similarity.

What would settle it

Independently reproduce the training procedure on the same BraTS-Local-Inpainting training data and submit the resulting model to the online validation platform; if the mean SSIM, PSNR, and MSE differ from 0.841, 23.257, and 0.007 by more than the paper's reported dispersions, the reported quantitative result does not reproduce.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a U-Net with three downsampling and three upsampling stages, skip connections, random healthy-mask augmentation, and a weighted MAE plus SSIM loss reconstructs healthy brain tissue in masked MRI volumes, achieving a validation SSIM of 0.841, a PSNR of 23.257, and an MSE of 0.007 as computed by the challenge's online evaluation platform. The model took first place in the BraTS-Local-Inpainting challenge. The authors trace the performance to the diversity created by five generated healthy masks per scan, augmented with random mirroring and rotation, and to the combination of a region-specific pixel error with a global structural similarity term.

Load-bearing premise

The challenge-provided healthy masks are assumed to correctly mark tissue that is independent of the tumor, so that training and evaluation on those masks genuinely measures how well the model synthesizes healthy brain tissue rather than how well it copies or adapts tumor-adjacent structures.

Editorial extensions

If this is right

The trained model can generate synthetic healthy counterparts of pathological MRI scans, which can be used as augmented training data for segmentation, registration, and other downstream models that need healthy references.
A standard U-Net with these simple additions establishes a strong, easily replicable baseline for the BraTS-Local-Inpainting task.
The relatively low standard deviations across validation metrics indicate the model performs consistently across diverse scans, not just on a handful of favorable cases.
The method requires only the challenge-provided masks and T1 images, so it can be applied directly to new BraTS-Local-Inpainting data without extra annotations.

Reading between the lines

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

If the healthy masks are not perfectly free of tumor-adjacent tissue, the reported scores may partly reflect reconstruction of residual abnormality; evaluating the same model with masks from a different generation algorithm would isolate this effect.
The MAE loss is a plausible cause of the blurriness the authors observe in low-intensity regions; replacing it with a perceptual loss or measuring MAE only within the healthy region might sharpen details at the cost of global similarity.
The SSIM term is computed on the whole image, so part of the optimization targets preserving the surrounding context rather than synthesizing the missing tissue; computing SSIM only inside the healthy mask would reveal how much of the score comes from preservation versus generation.
The evaluation uses only T1 MRI from the BraTS distribution; extending the same training recipe to other modalities or to out-of-distribution tumor types would test the method's generalization beyond the challenge.

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.

Desk Editor's Note

read the letter

Quick take: this is a straightforward U-Net entry into the BraTS 2023 inpainting challenge, and the paper is honest about that. The architecture is standard (U-Net with PReLU, instance norm, dropout), and the only real methodological choices are random mask augmentation and an MAE+SSIM combined loss. The evaluation numbers on the validation set were computed by the Synapse platform, which is independent, and the authors report metrics with standard deviations and show best/median/worst examples. That is more than many challenge papers do.

What is actually new: almost nothing. No new architecture, no new loss, no theoretical contribution. The contribution is the specific configuration and its reported performance on the challenge. The 'first place' claim is the key comparative evidence, and it is unsupported: no leaderboard screenshot, submission ID, or team name. That is a problem because without it there are no baselines at all. Absolute SSIM/PSNR values do not tell you whether the method is competitive. The internal inconsistency between the abstract (PSNR 23.257, MSE 0.007) and Table 1 (PSNR 23.381, MSE 0.0065) is minor, likely rounding or a different run, but it adds a bit of doubt about which numbers are official.

The paper also has a few small issues: Section 2.2 mentions 'BraTS 2022' instead of 2023, the SSIM formula is mislabeled, and the claim that the healthy masks 'accurately delineated' relies entirely on the challenge's mask generation algorithm. None of these are load-bearing. The method is standard enough that it probably works as described, and the external evaluation is a real check.

Who this is for: someone looking for a simple baseline for brain inpainting, or a participant in the BraTS inpainting challenge. As a standalone paper, it is thin. I would accept it for peer review only if the venue is a workshop or challenge-proceedings format, and I would make the first-place claim verifiable plus add at least one baseline (e.g., a vanilla U-Net without random masks). Without that, it is a tech report.

Referee Report

3 major / 5 minor

Summary. The paper proposes a 3D U-Net architecture for the ASNR-MICCAI BraTS Local Synthesis of Tissue via Inpainting task. The model takes a T1-weighted MRI scan with healthy and unhealthy regions masked out and predicts the missing healthy tissue. Training uses random generation of five healthy masks per scan, random mirroring and rotation, and a combined MAE-plus-SSIM loss. Validation metrics on the BraTS-Local-Inpainting validation set are reported as SSIM 0.841, PSNR 23.257, and MSE 0.007, with the statement that the method secured first place in the challenge. The paper concludes that the U-Net configuration reconstructs healthy brain tissue effectively.

Significance. If the reported metrics hold, this is a straightforward but solid application of a standard U-Net with random-mask augmentation to a clinically motivated inpainting task. The use of an external evaluation platform (Synapse) is a genuine strength, as are the 5-fold cross-validation and the detailed analysis of best/median/worst cases. However, the paper contributes no new architecture, no baseline comparison, and no ablation study. The central claim of effectiveness therefore rests almost entirely on the unverified 'first place' assertion. With the internal numerical inconsistencies, the evidence as presented is insufficient to support the paper's headline claims, but the issues are fixable within the scope of the manuscript.

major comments (3)

[Abstract and Table 1] The abstract and the conclusion report PSNR 23.257 (std 4.213), MSE 0.007 (std 0.007), and SSIM 0.841 (std 0.103), whereas Table 1 reports a mean PSNR of 23.3814246 (std 4.2644961), a mean MSE of 0.006503617 (std 0.004660640), and a mean SSIM of 0.8411663 (std 0.1031785). These differences are too large to be rounding artifacts, and the MSE standard deviation differs by a factor of 1.5. This inconsistency casts doubt on the provenance of the headline numbers. Please reconcile the reported values and state explicitly which numbers were computed by the Synapse platform.
[Abstract and Section 4] The claim that 'our method also secured first place in the challenge' appears in the Abstract and is repeated in the conclusion, but the paper provides no leaderboard link, team name, submission identifier, or challenge report. This is the only comparative evidence in the manuscript. Without a verifiable source, the claim should either be substantiated with a citation or removed. If removed, the absolute SSIM/PSNR/MSE values alone do not establish that the method is 'effective' relative to other approaches.
[Section 3.3 and Section 2.4] No baseline or ablation experiments are reported. The paper attributes success to random mask augmentation, five masks per scan, and the MAE+SSIM loss, but it never tests a U-Net trained with, for example, MAE loss alone or without the random-mask augmentation. The assertion in Section 2.4 that 'SSIM loss significantly underperforms MAE loss' is unsupported by any experiment. Adding at least one simple baseline and one loss ablation would make the central claim testable and would also help justify the design choices.

minor comments (5)

[Section 2.2] The preprocessing text says 'The BraTS 2022 dataset underwent standard pre-processing steps,' but the paper uses the BraTS-Local-Inpainting 2023 dataset. This is likely a typo, but it should be corrected to refer to the correct dataset.
[Section 2.1] The text says the training set contains 'four distinct data types' but then lists five bullet items: t1n, t1n-voided, healthy mask, unhealthy mask, and mask. Either the count or the list should be corrected.
[Section 3.1] The sentence 'It is worth noting that we only the healthy regions defined by healthy masks against the ground truth data' is missing a verb. It should read 'we only evaluate the healthy regions...'.
[Equation (3)] The paper calls SSIM a loss function but uses the similarity index directly in Equation (2). If the model minimizes the combined loss, a standard practice is to use (1 - SSIM) for the SSIM term. Please clarify whether the sign convention is intentional.
[Section 2.4 and Figure 1] The network input is described as 't1n-voided images and masks,' but the figure and text do not specify the number of input channels or how the mask is concatenated with the image. This information is needed for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the U-Net is trained and evaluated against external ground-truth healthy tissue via an independent challenge platform.

full rationale

The paper's derivation chain is a standard supervised image-inpainting pipeline. The target (healthy tissue) is defined by the external BraTS-Local-Inpainting challenge mask generation algorithm of Kofler et al. [2], not by the proposed model. Training uses pairs of masked inputs (t1n-voided) and ground-truth t1n volumes, with the loss in Eq. (3) comparing model output to ground truth via MAE and SSIM. Validation metrics (Table 1) were computed by the Sage Bionetworks Synapse platform on a held-out validation set, and only healthy-region metrics were used. The reported SSIM/PSNR/MSE therefore measure agreement between predicted and actual healthy tissue against an external benchmark, not against any quantity defined by the model or by a fitted parameter. No self-citation is load-bearing: references [1]–[4] are external datasets and standard architecture/metrics sources. The unsupported 'first place' claim and the numeric inconsistencies between the abstract/conclusion and Table 1 are verification concerns, not circularity; they do not show that any prediction reduces by construction to its input. Hence no circular step is present.

Assumptions & free parameters 5 free parameters · 2 assumptions · 0 invented entities

The central claim rests on the accuracy of the challenge's healthy masks and the representativeness of the BraTS dataset. The only hand-set values are standard hyperparameters, not fitted to the target result.

free parameters (5)

loss weight lambda_1 = 1
Hand-chosen in Section 3.2 for the MAE term of the combined loss.
loss weight lambda_2 = 1
Hand-chosen in Section 3.2 for the SSIM term of the combined loss.
learning rate = 1e-4
Adam optimizer initial learning rate set in Section 3.2.
dropout rate = 0.2
Set in Section 3.2 for the bridge and upsampling blocks.
number of masks per scan = 5
Healthy mask generation count described in Section 2.3.

assumptions (2)

domain assumption The healthy mask generation algorithm in [2] accurately delineates healthy brain regions that are spatially separated from the tumor.
Section 2.1 relies on this to define training targets and evaluation regions.
domain assumption The BraTS-Local-Inpainting dataset is representative of brain MRI scans with tumors.
Used as the sole training and evaluation data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of U-Net Based Healthy 3D Brain Tissue Inpainting." pith.science (2026). https://pith.science/paper/SPZGBIZ6

@misc{pith2026250718126,
  author       = {Pith},
  title        = {Pith review of: U-Net Based Healthy 3D Brain Tissue Inpainting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SPZGBIZ6}},
  note         = {Machine review of arXiv:2507.18126}
}

read the original abstract

This paper introduces a novel approach to synthesize healthy 3D brain tissue from masked input images, specifically focusing on the task of 'ASNR-MICCAI BraTS Local Synthesis of Tissue via Inpainting'. Our proposed method employs a U-Net-based architecture, which is designed to effectively reconstruct the missing or corrupted regions of brain MRI scans. To enhance our model's generalization capabilities and robustness, we implement a comprehensive data augmentation strategy that involves randomly masking healthy images during training. Our model is trained on the BraTS-Local-Inpainting dataset and demonstrates the exceptional performance in recovering healthy brain tissue. The evaluation metrics employed, including Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Mean Squared Error (MSE), consistently yields impressive results. On the BraTS-Local-Inpainting validation set, our model achieved an SSIM score of 0.841, a PSNR score of 23.257, and an MSE score of 0.007. Notably, these evaluation metrics exhibit relatively low standard deviations, i.e., 0.103 for SSIM score, 4.213 for PSNR score and 0.007 for MSE score, which indicates that our model's reliability and consistency across various input scenarios. Our method also secured first place in the challenge.

Figures

Figures reproduced from arXiv: 2507.18126 by the authors.

**Figure 1.** Architecture of our U-Net model. It is important to note that our model processes cropped images as both input and output, which differs slightly from the visualization in the figure. Typically, these cropped images encompass the entire brain tissue. 2.3 Data Augmentation Deep learning models with extensive parameters often suffer from overfitting, hindering their ability to generalize to unseen data. To mitigate … view at source ↗

**Figure 2.** Visualization of the infilled validation MRI scans. In the images, the orange portions represent the masked regions containing both healthy and unhealthy tissues, since they were not explicitly labeled in the validation dataset. During the validation phase, our model’s performance has been exclusively assessed using healthy tissue segments [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

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

Now You Have My Healthy Attention: A U-DiT for Brain-MRI Inpainting
cs.CV 2026-07 accept novelty 5.5 of 10

Healthy-only bottleneck attention plus a contralateral-mirror input raise BraTS-2026 healthy-tissue inpainting to SSIM 0.864 / PSNR 24.7 dB on 219 validation cases.

Reference graph

Works this paper leans on

4 extracted references · 1 canonical work pages · cited by 1 Pith paper

[1]

Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The rsna-asnr-miccai brats 2021benchmarkonbraintumorsegmentationandradiogenomicclassification.arXiv preprint arXiv:2107.02314 (2021)

arXiv 2021
[2]

arXiv preprint arXiv:2305.08992 (2023)

Kofler, F., Meissen, F., Steinbauer, F., Graf, R., Oswald, E., de da Rosa, E., Li, H.B., Baid, U., Hoelzl, F., Turgut, O., et al.: The brain tumor segmentation (brats) challenge 2023: Local synthesis of healthy brain tissue via inpainting. arXiv preprint arXiv:2305.08992 (2023)

arXiv 2023
[3]

In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18. pp. 234–241. Springer (2015)

2015
[4]

IEEE Transactions on Image Processing 13(4), 600–612 (2004)

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

work page 2004

Pith tools

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

[1] [1]

Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The rsna-asnr-miccai brats 2021benchmarkonbraintumorsegmentationandradiogenomicclassification.arXiv preprint arXiv:2107.02314 (2021)

arXiv 2021

[2] [2]

arXiv preprint arXiv:2305.08992 (2023)

Kofler, F., Meissen, F., Steinbauer, F., Graf, R., Oswald, E., de da Rosa, E., Li, H.B., Baid, U., Hoelzl, F., Turgut, O., et al.: The brain tumor segmentation (brats) challenge 2023: Local synthesis of healthy brain tissue via inpainting. arXiv preprint arXiv:2305.08992 (2023)

arXiv 2023

[3] [3]

In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18. pp. 234–241. Springer (2015)

2015

[4] [4]

IEEE Transactions on Image Processing 13(4), 600–612 (2004)

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

work page 2004