SynBT: High-quality Tumor Synthesis for Breast Tumor Segmentation by 3D Diffusion Model

Edina Timko; Hongxu Yang; Lehel Ferenczi; Levente Lippenszky; Vanda Czipczer

REVIEW 3 major objections 6 minor 14 references

SynBT: High-quality Tumor Synthesis for Breast Tumor Segmentation by 3D Diffusion Model

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Adding SynBT synthetic breast tumors to training data improves MRI tumor segmentation Dice by 2-3 points across three architectures.

desk verdict Consistent but unverified 2-3% Dice gains from a sensible large-volume synthetic tumor pipeline; the lack of error bars is the main thing to fix before trusting the headline. read the letter →

arxiv 2509.03267 v1 pith:UEEHTH7S submitted 2025-09-03 cs.CV

Hongxu Yang , Edina Timko , Levente Lippenszky , Vanda Czipczer , Lehel Ferenczi This is my paper · ORCID

classification cs.CV

keywords synthetictumorgenerationdiffusionmodelVQ-VAEbreastsegmentationDCE-MRIlatentspaceinpaintingdataaugmentation3Dmedicalimagesynthesis

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 claims that realistic synthetic breast tumors, generated in full-volume contrast-enhanced MRI by a mask-conditioned 3D diffusion model, can be used as supplementary training data to improve tumor segmentation. It argues that previous tumor synthesis, designed for small patches, fails when tumors sit in large field-of-view volumes like breast MRI. SynBT generates tumors directly in latent space after compressing high-resolution MRIs with a patch-to-volume VQ-VAE, then places the synthetic tumors near fibroglandular tissue using masks from a template bank. On a large public DCE-MRI dataset, training U-Net, nnU-Net, and SwinUNETR with SynBT data improved Dice by roughly 2-3% over training on real volumes alone. The authors present this as evidence that controllable foreground synthesis, rather than full-image generation, is a practical path for boosting segmentation when annotated tumors are scarce.

What carries the argument

The load-bearing mechanism is the combination of a patch-to-volume VQ-VAE with a mask-conditioned latent diffusion model. The VQ-VAE first trains on small patches to learn local representations, then reassembles patched latents into a full-volume latent and fine-tunes only the decoder, avoiding grid artifacts while keeping the large field of view. The diffusion model works in this latent space, inpainting tumor texture into a masked region conditioned on the downsampled binary mask; at inference, a template-bank mask is adjusted by fibroglandular-tissue placement rules to control tumor shape, size, and location. This design lets the generator model only the foreground tumor, not the whole br

What would settle it

Retrain the best nnU-Net on SynBT data where synthetic tumor masks are placed randomly in breast tissue instead of aligned to fibroglandular tissue; if the Dice gain over the real-volume baseline persists (at least 2 points), the placement rule is not the mechanism. Alternatively, segment a test cohort whose tumors are predominantly in fatty tissue, where the reported improvement should shrink if the assumption is load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that the proposed SynBT pipeline produces synthetic breast tumors realistic enough to serve as training data, and that they improve segmentation performance consistently. The method compresses paired pre-/post-contrast MRI volumes into a compact latent space using a two-step patch-to-volume VQ-VAE, trains a diffusion model to inpaint tumor textures conditioned on a binary mask, and then generates on-the-fly synthetic tumors by sampling masks from a template bank and placing them near fibroglandular tissue. The paper reports Dice Score improvements from 0.7463 to 0.7629 for nnU-Net, 0.7278 to 0.7580 for SwinUNETR, and 0.7220 to 0.7326 for U-Net, alongside reduced Hausdorf

Load-bearing premise

Tumors in the test set appear in breast tissue in roughly the same way as the training-derived template bank and fibroglandular placement rule assume; if real tumors often sit outside fibroglandular tissue, the synthetic data could mislead the segmentation model.

Editorial extensions

If this is right

Training with SynBT data improves Dice, 95% Hausdorff distance, and mean surface distance across three different segmentation architectures, suggesting the benefit is not tied to one model.
Because the generator only synthesizes tumor foreground on healthy backgrounds, it can reuse unannotated or healthy breast MRIs, lowering the annotation burden for new segmentation tasks.
On-the-fly sampling of masks from a template bank gives controllable variety in tumor size and location during segmentation training.
Patch-based training of the VQ-VAE keeps memory practical for high-resolution volumes, making latent diffusion feasible for large-FOV 3D images.
The 2-3% Dice gain on a public benchmark gives a concrete reference point for comparing future tumor synthesis methods.

Reading between the lines

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

Because the framework models only tumor texture and receives the background from real MRIs, its central components could transfer to other contrast-enhanced tumor types with large FOV, such as liver or prostate MRI, although the paper does not test this.
The paper evaluates quality only through downstream segmentation; a direct test of whether synthetic tumors match real radiomics features (FID/SSIM/radiomics similarity) remains open.
A sharper placement model, using clinical priors rather than a fibroglandular-tissue heuristic, could improve both realism and the resulting segmentation gain, as the authors acknowledge.
One could test the contribution of each component by ablating the template bank or the tissue-placement rule; the paper does not isolate these, so their individual weights are unknown.

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

What you should know: this is a solid engineering paper that reports a consistent but modest improvement (about 1-3 Dice points) from synthetic breast tumors for segmentation on the public MAMA-MIA dataset. The main risk is that the gain rests on single runs with no variance estimates, so the headline number could be seed noise rather than a real effect.

What's new: the patch-to-volume VQ-VAE with multi-sequence input and decoder-only volume fine-tuning is a reasonable adaptation of 3D MedDiffusion, and applying mask-conditioned latent diffusion to whole large-FOV breast DCE-MRI volumes appears to be new territory. The paper also shows something useful: patch-based segmentation fails badly on this data, and the synthetic data helps all three volume-based models (U-Net, nnU-Net, SwinUNETR). That consistency across architectures is the best evidence for the method. The comparison against SynTumor is meaningful, though SynTumor was designed for 64^3 patches, so it is at a disadvantage in this large-volume setting; the authors acknowledge this.

The soft spots are real but not fatal. The stress-test is right: Table 1 has no error bars, confidence intervals, or repeated runs. The effect size is small enough that run-to-run variance could explain it. The paper itself acknowledges there is no quantitative image-quality evaluation (FID, SSIM, radiomics) and that tumor placement rests on an assumption that tumors generally sit at the fibroglandular tissue. Both are genuine limitations. There is also no code or data release, which makes reproduction harder.

Who this is for: people working on tumor synthesis or breast MRI segmentation who want a credible recipe for large-volume generative augmentation. It deserves a serious referee. The engineering is sound and the result is plausible enough to warrant independent verification, but the authors should be asked for error bars, code, and at least one quantitative image-quality metric before the claim of 2-3% improvement is accepted as established.

Referee Report

3 major / 6 minor

Summary. The paper proposes SynBT, a 3D latent diffusion framework for synthesizing breast tumors in large-field-of-view DCE-MRI volumes. A two-step patch-to-volume VQ-VAE compresses dual-sequence volumes into a compact latent space; a mask-conditioned diffusion model then inpaints tumor texture in this latent space, and tumor placement is guided by a template bank of real tumor masks and fibroglandular tissue segmentation. The synthetic volumes are used on-the-fly as supplementary training data for three segmentation architectures (U-Net, nnU-Net, SwinUNETR). Evaluated on the MAMA-MIA public dataset, SynBT-augmented training is reported to improve Dice over volume-based training by 1.1, 1.7, and 3.0 percentage points for the three architectures, respectively (Table 1). The paper also compares against SynTumor and reports that this baseline performs worse.

Significance. If the reported segmentation gains are statistically reliable, SynBT is a meaningful contribution to tumor synthesis in large-FOV 3D MRI. The paper is the first to validate whole-volume synthetic breast-tumor augmentation on a large public benchmark, across three different segmentation models, with consistent directional improvements in Dice, 95HD, and MSD. The use of a held-out test split with real ground-truth masks avoids circularity. The method is practical in that it combines a tumor template bank with tissue-based placement, and the pipeline is described in sufficient detail for replication. However, the central quantitative claim currently rests on single training runs per configuration, and no quantitative image-quality metrics are provided, so the strength of the evidence is conditional on additional statistical validation.

major comments (3)

[Table 1, §3.1–3.2] The central claim of a 2–3% Dice improvement is supported only by single training runs for each configuration. With 206 test volumes, differences of 1.1, 1.7, and 3.0 percentage points could plausibly fall within seed-to-seed variance, especially because models are trained for a fixed 300 epochs and selected by validation score. Please report mean ± std over at least 3–5 random seeds and provide a paired significance test (e.g., Wilcoxon signed-rank or bootstrap) for DSC, 95HD, and MSD. The abstract should also state the actual observed range (U-Net +1.1, nnU-Net +1.7, SwinUNETR +3.0) rather than the aggregate '2–3%'.
[§3.2, SynTumor baseline] The SynTumor comparison likely places that method outside its intended operating regime: SynTumor was designed for small-patch (e.g., 96^3) CT tumor synthesis, while here it is applied to full 128×192×192 MRI volumes with large FOV. This may unfairly handicap the baseline and weakens the claim that SynBT is superior to the state of the art. Please either adapt SynTumor to the large-FOV setting (e.g., patch-based generation followed by stitching) or limit the comparison claim to 'our volume-based pipeline with SynBT versus an out-of-the-box SynTumor baseline.'
[§4, Fig. 4] The title and abstract describe the synthetic tumors as 'high-quality,' but no quantitative image-quality metric is reported. Fig. 4 provides only qualitative visual examples, and the paper itself (Sec. 4) defers FID, SSIM, and radiomics-based evaluation to future work. Because the segmentation improvement is the primary evidence of utility, please either add quantitative synthesis-quality metrics or temper the 'high-quality' characterization to 'visually realistic in qualitative inspection.'

minor comments (6)

[§1] The framework is called 'SynthBT' in the Introduction but 'SynBT' in the title and abstract. Please unify the name.
[§2.1] 'of size963 voxels' appears to be a typo for 'of size 96^3 voxels.' Please fix.
[§3.1] 'isotopically resampled' should be 'isotropically resampled.'
[§3.1] Please specify the number of synthetic volumes generated per training batch/epoch and the ratio of synthetic to real samples, as this is key for reproducibility.
[§2.2] The conditioning input Zcond = cat(Zmasked, down(M)) is not fully specified; please state the concatenation axis and the exact down-sampling operation used for the mask.
[§3.2] The patch-based segmentation results (DSC ≈ 0.38–0.47) are dramatically lower than volume-based results. A brief explanation of why patch-based methods fail so severely in this high-FOV DCE-MRI setting would help readers interpret the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the segmentation evaluation is held out and the synthesis pipeline is trained only on training data.

full rationale

The paper's central claim is empirical: synthetic tumors generated by SynBT improve breast tumor segmentation Dice by 1-3 percentage points over training on real volumes alone. The evaluation uses the official MAMA-MIA split with a held-out test set of 206 volumes; the synthesis models (VQ-VAE, latent diffusion, template bank, fibroglandular tissue placement rule) are all derived from training data and are never re-fit or tuned on the test set. No equation in the paper reduces a prediction to its own input, and no fitted parameter is renamed as a prediction. The method's assumptions, explicitly stated in the Conclusions (tumors generally exist at randomized locations in fibroglandular tissue; future FID/SSIM/Radiomics quality analysis is needed), are limitations on realism and external validity, not circularities, because the test-set segmentation masks are real ground truth and the reported improvement is not forced by construction. The absence of error bars and significance tests is a correctness/statistical robustness concern, but it does not make the derivation circular. The paper also does not rely on load-bearing self-citations: the cited prior works (VQ-VAE, DDPM, MedDiffusion, SynthSeg, SynTumor, etc.) are external and are used as components or baselines, not to define the claimed result.

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

The paper is entirely empirical and introduces no new physical or conceptual entities. The central claim, improved segmentation from synthetic data, rests on the validity of the synthetic data generation pipeline, which relies on assumptions about latent space fidelity, template bank representativeness, tissue model transferability, and contrast-phase consistency. Free parameters are standard hyperparameters chosen by hand rather than fitted to the test metric.

free parameters (5)

VQ-VAE codebook size K = 8192
Hand-chosen; governs latent representational capacity and impacts synthetic tumor fidelity, but is not fitted to the test metric.
VQ-VAE compression rate = 4
Hand-chosen trade-off between memory usage and reconstruction quality.
Patch training size = 64^3
Patch size for the first VQ-VAE training step; influences the local context learned.
Diffusion latent crop size = 32x48x48x8
Latent dimensions of the sliding-window crops used for diffusion and segmentation; set to fit GPU memory.
Tumor placement post-processing thresholds = unspecified
Rules selecting 'the most possible regions' for tumor placement are not quantified in the paper, directly affecting reproducibility.

assumptions (5)

domain assumption The VQ-VAE latent space preserves tumor texture sufficiently for mask-conditioned diffusion to synthesize realistic tumors.
Sections 2.1 and 2.2 assume a codebook size of 8192 and compression rate 4 retain clinically relevant appearance; no reconstruction quality metric is reported.
domain assumption Tumor masks in the template bank, taken from the MAMA-MIA training set, are representative of tumor shape, size, and location in the test set.
Section 2.3 builds the bank from training tumors; if the test distribution differs, synthetic augmentation may not transfer.
domain assumption The fibroglandular tissue model trained on Duke DCE-MRI with SynthSeg generalizes to MAMA-MIA images.
Section 3.1 uses Duke fibroglandular masks to train the tissue segmenter, then applies it to MAMA-MIA for tumor placement; domain shift is not quantified.
domain assumption Selecting a random post-contrast sequence during training is equivalent to using the first post-contrast sequence at test time.
Section 3.1 states this choice; if enhancement timing differs, the synthesis may not match the test distribution.
domain assumption The diffusion model generates tumor texture indistinguishable from real enhancing tumor tissue.
Section 2.2 and Figure 4 provide qualitative evidence but no quantitative metrics; the segmentation improvement is used as indirect evidence.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SynBT: High-quality Tumor Synthesis for Breast Tumor Segmentation by 3D Diffusion Model." pith.science (2026). https://pith.science/paper/UEEHTH7S

@misc{pith2026250903267,
  author       = {Pith},
  title        = {Pith review of: SynBT: High-quality Tumor Synthesis for Breast Tumor Segmentation by 3D Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UEEHTH7S}},
  note         = {Machine review of arXiv:2509.03267}
}

read the original abstract

Synthetic tumors in medical images offer controllable characteristics that facilitate the training of machine learning models, leading to an improved segmentation performance. However, the existing methods of tumor synthesis yield suboptimal performances when tumor occupies a large spatial volume, such as breast tumor segmentation in MRI with a large field-of-view (FOV), while commonly used tumor generation methods are based on small patches. In this paper, we propose a 3D medical diffusion model, called SynBT, to generate high-quality breast tumor (BT) in contrast-enhanced MRI images. The proposed model consists of a patch-to-volume autoencoder, which is able to compress the high-resolution MRIs into compact latent space, while preserving the resolution of volumes with large FOV. Using the obtained latent space feature vector, a mask-conditioned diffusion model is used to synthesize breast tumors within selected regions of breast tissue, resulting in realistic tumor appearances. We evaluated the proposed method for a tumor segmentation task, which demonstrated the proposed high-quality tumor synthesis method can facilitate the common segmentation models with performance improvement of 2-3% Dice Score on a large public dataset, and therefore provides benefits for tumor segmentation in MRI images.

Figures

Figures reproduced from arXiv: 2509.03267 by the authors.

**Figure 1.** Patch-to-Volume VQ-VAE training has two steps. First, the VQ-VAE is trained on random sub-patch from the input MRI images (two sequences with one pre-contrast MRI and a random post-contrast MR image). The VQ-VAE is trained to compress and reconstruct the input information in the latent space with a quantization codebook. Second, the input volumes are divided into small patches, which are processed by patchlevel enc… view at source ↗

**Figure 2.** With trained Patch-to-Volume VQ-VAE, a latent space diffusion model is trained for synthetic tumor generation. In this stage, tumors in the volumes are masked by ground truth, which are processed by the encoder of VQ-VAE to generate latent space vectors without tumor. Tumors are synthetically generated by diffusion model from the conditioned binary mask. The final stage of the framework is the training of the segmen… view at source ↗

**Figure 3.** The synthetic tumor generation. With given input pre- and post-contrast MRIs, the tumor mask template is selected from the tumor template bank. The tumor mask template is processed to generate final tumor mask by combining with ROI selection and fibroglandular tissue segmentation, which are controlling the size, position, shape, etc. of the tumor. Finally, the synthetic tumor is generated in pre-/post-contrast MRIs.… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: (a) Left to right: real image, patch-only VQ-VAE output and Patch-to-Volume VQ-VAE outputs. Top row is post-contrast MRI, while bottom row is pre-contrast MRI. Grid effect can be observed in mid-column images for patch-only VQ-VAE. (b) SynthSeg based method for fibrogl…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 9 canonical work pages

[1]

Medical image analysis86, 102789 (2023)

Billot, B., Greve, D.N., Puonti, O., Thielscher, A., Van Leemput, K., Fischl, B., et al.: Synthseg: Segmentation of brain mri scans of any contrast and resolution without retraining. Medical image analysis86, 102789 (2023)

2023
[2]

arXiv preprint arXiv:2211.02701 (2022)

Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., et al.: Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)

arXiv 2022
[3]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Chen, Q., Chen, X., Song, H., Xiong, Z., Yuille, A., Wei, C., et al.: Towards general- izable tumor synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11147–11158 (2024)

work page 2024
[4]

Nature Biomedical Engineering 5(6), 493–497 (2021)

Chen, R.J., Lu, M.Y., Chen, T.Y., Williamson, D.F., Mahmood, F.: Synthetic data in machine learning for medicine and healthcare. Nature Biomedical Engineering 5(6), 493–497 (2021)

work page 2021
[5]

arXiv e-prints pp

Garrucho, L., Reidel, C.A., Kushibar, K., Joshi, S., Osuala, R., Tsirikoglou, A., et al.: Mama-mia: A large-scale multi-center breast cancer dce-mri benchmark dataset with expert segmentations. arXiv e-prints pp. arXiv–2406 (2024)

work page 2024
[6]

In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

Guo, P., Zhao, C., Yang, D., Xu, Z., Nath, V., Tang, Y., et al.: Maisi: Medical ai for synthetic imaging. In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4430–4441. IEEE (2025)

work page 2025
[7]

Advances in neural information processing systems33, 6840–6851 (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

2020
[8]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Hu, Q., Chen, Y., Xiao, J., Sun, S., Chen, J., Yuille, A.L., et al.: Label-free liver tumor segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7422–7432 (2023)

work page 2023

Show all 14 references

[9]

Scientific Reports13(1), 7303 (2023)

Khader, F., Müller-Franzes, G., Tayebi Arasteh, S., Han, T., Haarburger, C., Schulze-Hagen, M., et al.: Denoising diffusion probabilistic models for 3d medi- cal image generation. Scientific Reports13(1), 7303 (2023)

2023
[10]

the cancer imaging archive

Saha, A., Harowicz, M., Grimm, L., Weng, J., Cain, E., Kim, C., et al.: Dynamic contrast-enhanced magnetic resonance images of breast cancer patients with tumor locations [data set]. the cancer imaging archive. Bethesda, MD, USA (2021) 10 H. Yang et al

2021
[11]

arXiv preprint arXiv:2010.02502 (2020)

Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

2010 arXiv
[12]

IEEE journal of biomedical and health informatics26(8), 3966–3975 (2022)

Sun, L., Chen, J., Xu, Y., Gong, M., Yu, K., Batmanghelich, K.: Hierarchical amortized gan for 3d high resolution medical image synthesis. IEEE journal of biomedical and health informatics26(8), 3966–3975 (2022)

2022
[13]

Advances in neural information processing systems30 (2017)

Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems30 (2017)

2017
[14]

IEEE Transactions on Medical Imaging (2025)

Wang, H., Liu, Z., Sun, K., Wang, X., Shen, D., Cui, Z.: 3d meddiffusion: A 3d medical latent diffusion model for controllable and high-quality medical image generation. IEEE Transactions on Medical Imaging (2025)

2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Medical image analysis86, 102789 (2023)

Billot, B., Greve, D.N., Puonti, O., Thielscher, A., Van Leemput, K., Fischl, B., et al.: Synthseg: Segmentation of brain mri scans of any contrast and resolution without retraining. Medical image analysis86, 102789 (2023)

2023

[2] [2]

arXiv preprint arXiv:2211.02701 (2022)

Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., et al.: Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)

arXiv 2022

[3] [3]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Chen, Q., Chen, X., Song, H., Xiong, Z., Yuille, A., Wei, C., et al.: Towards general- izable tumor synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11147–11158 (2024)

work page 2024

[4] [4]

Nature Biomedical Engineering 5(6), 493–497 (2021)

Chen, R.J., Lu, M.Y., Chen, T.Y., Williamson, D.F., Mahmood, F.: Synthetic data in machine learning for medicine and healthcare. Nature Biomedical Engineering 5(6), 493–497 (2021)

work page 2021

[5] [5]

arXiv e-prints pp

Garrucho, L., Reidel, C.A., Kushibar, K., Joshi, S., Osuala, R., Tsirikoglou, A., et al.: Mama-mia: A large-scale multi-center breast cancer dce-mri benchmark dataset with expert segmentations. arXiv e-prints pp. arXiv–2406 (2024)

work page 2024

[6] [6]

In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

Guo, P., Zhao, C., Yang, D., Xu, Z., Nath, V., Tang, Y., et al.: Maisi: Medical ai for synthetic imaging. In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4430–4441. IEEE (2025)

work page 2025

[7] [7]

Advances in neural information processing systems33, 6840–6851 (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

2020

[8] [8]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Hu, Q., Chen, Y., Xiao, J., Sun, S., Chen, J., Yuille, A.L., et al.: Label-free liver tumor segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7422–7432 (2023)

work page 2023