Masked Autoencoder Pretraining and BiXLSTM ResNet Architecture for PET/CT Tumor Segmentation

Abdul Qayyum; Moona Mazher; Steven A Niederer

REVIEW 3 major objections 4 minor 14 references

Masked Autoencoder Pretraining and BiXLSTM ResNet Architecture for PET/CT Tumor Segmentation

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that self-supervised masked-autoencoder pretraining raises mean Dice from 0.543 to 0.582 for whole-body PET/CT tumor segmentation on the AutoPET Task 1 dataset.

desk verdict MAE pretraining adds roughly 0.04 Dice for this PET/CT segmentation model—plausible, but the paper's 'significant' is unsupported by any variance or significance test. read the letter →

arxiv 2509.02602 v1 pith:Z4TPHJDW submitted 2025-08-29 eess.IV

Moona Mazher , Steven A Niederer , Abdul Qayyum This is my paper · ORCID

classification eess.IV

keywords PET/CTlesionsegmentationmaskedautoencoderself-supervisedpretrainingBiXLSTMResNetmultimodalfusionAutoPETdatasettumor

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 tries to establish that a two-stage pipeline—masked-autoencoder pretraining on unlabeled PET/CT plus longitudinal CT scans, followed by fine-tuning a bidirectional XLSTM–ResNet encoder with a convolutional decoder—improves whole-body tumor segmentation on the AutoPET Task 1 benchmark. It reports a mean Dice gain of 0.039 (0.543 from scratch to 0.582 with pretraining) and a reduction in false-negative volume from 23.26 to 15.08 cm³. This matters because labeled medical imaging is scarce and expensive, so the claim is that easily collected unlabeled scans can substitute for some annotation effort. The paper also argues that fusing PET and CT input channels helps the model combine metabolic and anatomical information. The key number is a single point estimate; the paper calls the improvement significant but presents no variance or significance test.

What carries the argument

The carrying mechanism is the two-stage transfer pipeline. Stage 1 uses a Masked Autoencoder (MAE), which masks patches of 3D PET/CT volumes and trains the encoder to reconstruct them, forcing it to encode global and local structure without labels. Stage 2 discards the decoder, keeps the pretrained encoder, and fine-tunes it as the backbone of a segmentation network whose other components are a bidirectional XLSTM (an extended LSTM reading slices in both directions), residual (ResNet) blocks, and a standard convolutional decoder; PET and CT volumes enter as fused input channels. The pretrained weights initialize the segmentation encoder, and the paper claims this initialization drives the Di

What would settle it

Run the same three configurations (no SSL, SSL on the encoder, SSL on the bottleneck) five times each with different random seeds, using the same 5-fold splits, and record the full Dice distribution. If the 0.039-point gap falls within one standard deviation of the two distributions, the claimed pretraining benefit is not distinguishable from training noise.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that self-supervised masked-autoencoder pretraining materially improves a PET/CT lesion segmentation model. Starting from the same BiXLSTM–ResNet UNet architecture, training from scratch reaches 0.543 mean Dice, while first pretraining the encoder with a masked autoencoder on unlabeled PET/CT and longitudinal CT scans and then fine-tuning with annotations reaches 0.582, along with lower false-negative volume and a small reduction in false-positive volume. The authors interpret this as evidence that the MAE phase learns robust modality-specific representations—anatomical structure from CT and metabolic activity from PET—that transfer to supervise

Load-bearing premise

The claimed benefit of self-supervised pretraining rests on a single comparison of mean Dice scores, 0.582 versus 0.543, with no reported variance across folds or training runs; if the two score distributions overlap, the gain could be noise.

Editorial extensions

If this is right

If the gain is real, self-supervised pretraining on unlabeled PET/CT and longitudinal CT scans can raise Dice by about four points without any additional manual annotation.
The reduced false-negative volume implies the pretrained model recovers more true lesions, which is the clinically more costly error in screening and staging.
The two SSL variants differ by only 0.002 Dice, suggesting that the main benefit comes from having done pretraining at all, not from exactly where the pretrained weights are attached.
The approach fits inside a standard self-configuring preprocessing and training pipeline, so it can be reproduced with moderate compute on a single GPU.
Because PET and CT are treated as complementary input channels, the framework is a concrete example of multimodal fusion benefiting a supervised downstream task.

Reading between the lines

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

The paper does not isolate whether the pretraining benefit comes from the PET/CT data, the longitudinal CT data, or both; an ablation pretraining on each source separately would localize the source of the gain.
The architecture is not inherently PET/CT-specific, so the same two-stage recipe is a natural starting point for other multimodal volumetric segmentation tasks, though the paper does not test transfer.
The 0.002 Dice gap between the encoder-attached and bottleneck-attached SSL variants suggests a follow-up question: where in the network do pretrained features matter most for lesion recovery?
Because Task 2 longitudinal CT data enters only through pretraining, this design hints that a related but differently annotated dataset can contribute to segmentation accuracy without extra labeling effort.

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 read: this is an AutoPET challenge write-up whose headline claim is that MAE self-supervised pretraining raises mean Dice from 0.543 to 0.582 for a BiXLSTM–ResNet inside nnU-Net. That direction is plausible and consistent with the broader SSL literature, and it's the paper's only real result. The problem: Table 2 gives means only—no standard deviations, no per-fold numbers, no significance test—yet the abstract and discussion use 'significantly.' The two SSL variants in the table sit 0.002 Dice apart (0.580 vs 0.582), so the unexplained noise band may well cover the 0.039 gain. The narrative also picks the better row for the headline; the other SSL row has lower false-negative volume but slightly lower Dice. That's selection in reporting, even if unintentional.

What's new: the specific combination of MAE pretraining on unlabeled PET/CT and longitudinal CT, followed by a BiXLSTM-ResNet encoder with PET/CT fusion, evaluated on AutoPET Task 1. It's an application of known components, not a new method. What's done well: the paper is honest about what went into the pipeline—nnU-Net preprocessing and augmentation, 5-fold CV, TTA, no external pretrained models, and using Task 2 data only for pretraining. The algorithm details table is useful.

Soft spots apart from the stats: the MAE recipe (mask ratio, patch size, epochs, data mix) is unreported, so others can't reproduce the pretraining; code is promised but not released. References 9 and 10 are identical duplicates. There's also no external leaderboard rank, which is odd for a challenge paper. These are all fixable.

Is the result real? Probably yes—0.04 Dice is about the size of gain you'd expect from SSL in this setting. But the paper does not demonstrate it. The right fix is straightforward: report per-fold results, standard deviations, a paired significance test across folds, and give the MAE configuration. If the gain holds, it's a modest but credible datapoint for the AutoPET community.

Verdict: worth a serious referee, and I'd want to see it back after those revisions. Not something I'd cite for method novelty, but I'd consider citing it as a challenge baseline if the stats get fixed.

Referee Report

3 major / 4 minor

Summary. The paper proposes a two-stage PET/CT lesion segmentation framework for the fourth AutoPET Challenge. In Stage 1, a Masked Autoencoder (MAE) is pretrained on unlabeled PET/CT and longitudinal CT data. In Stage 2, the pretrained encoder is integrated into a BiXLSTM–ResNet architecture with a convolutional decoder and fine-tuned on annotated AutoPET Task 1 data using a Dice-plus-cross-entropy loss, with PET and CT fused as input channels. On the AutoPET Task 1 benchmark, the authors report a mean Dice score of 0.582 with SSL pretraining versus 0.543 without, together with false-negative and false-positive volumes, and conclude that self-supervised pretraining significantly improves segmentation accuracy. The paper includes qualitative visualizations and states that code will be made available.

Significance. If the claimed improvement is reproducible, the paper would provide a useful evidence point for self-supervised pretraining in multimodal PET/CT segmentation, a clinically relevant task. The work is grounded in a public benchmark (AutoPET), uses the standard nnUNet v2 framework, and reports a two-stage architecture that is clearly described at a high level. However, the central quantitative claim rests entirely on a single comparison of mean Dice scores with no measure of variance, no per-fold results, and no significance test. Given the small absolute differences involved, the main conclusion is currently not statistically supported. The reader's stress-test concern is well-founded and is the primary obstacle to acceptance.

major comments (3)

[Section 3, Table 2] The central claim that SSL 'significantly enhances' segmentation (Abstract; Section 3; Section 4) is supported only by mean Dice values: 0.543 (no SSL), 0.580 (SSL Enc), 0.582 (SSL Bot). No standard deviations, per-fold numbers, or significance tests are reported. The two SSL variants differ by 0.002 Dice, which suggests that small architectural or training-stochasticity differences can produce variation on this scale; the reported SSL gain of 0.039 is only about twenty times that internal spread. Without error bars, a paired test, or repeated seeds, the 0.039 gain cannot be distinguished from run-to-run noise. Please report per-fold Dice, standard deviations or confidence intervals, and a paired significance test (e.g., Wilcoxon signed-rank over folds or a bootstrap) comparing SSL versus no-SSL.
[Section 3, Table 2] The narrative in Section 3 selects the BiXLSTM–ResNet_Bot + SSL row (Dice 0.582) for the headline result, while the other SSL row (BiXLSTM–ResNet_Enc + SSL) achieves a lower Dice (0.580) but a better false-negative volume (13.78 vs 15.08 cm³). The choice of the Bot row as the 'winner' is not justified by a pre-specified selection rule, and reporting one row while downplaying the other introduces possible selection bias. Both SSL variants should be reported transparently in the abstract and discussion, or the selection criterion should be stated.
[Section 2.4, Table 1] The Stage 1 MAE pretraining recipe is not specified: mask ratio, patch size, number of pretraining epochs, optimizer, learning rate, and the exact mix of unlabeled PET/CT and longitudinal CT data are all absent. Since the paper's contribution is the two-stage SSL pipeline, the pretraining configuration is load-bearing for reproducibility and for attributing the Dice gain to SSL rather than to a particular pretraining hyperparameter choice. Please provide these details or cite a specific configuration.

minor comments (4)

[References] References [9] and [10] are identical (Qayyum et al., 'Transforming Heart Chamber Imaging...'). Please remove the duplicate.
[Figures 2 and 3] The qualitative figures show only SSL-pretrained models; they do not include a from-scratch baseline. The claim that 'these figures confirm that self-supervised pretraining improves...' is therefore not supported by the figures themselves.
[Table 1 / Section 2.3] Some labels are inconsistent: 'BiXLSTM–ResNet_Bot + SSL' is used in Table 2 but the architecture description in Section 2.4 does not explain what 'Bot' refers to. Please define this variant.
[Section 2.3] The paper says 'nnUNet-compatible NIfTI formatting, including training/validation splits from splits_final.json' but does not state how many folds were used for the final test evaluation; '5-fold cross-validation' is mentioned later, but the mapping between folds and the reported metrics should be explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central SSL-versus-scratch comparison is an internal empirical ablation measured against the external AutoPET benchmark; self-citations are programmatic, not load-bearing.

full rationale

The paper's central claim is that MAE self-supervised pretraining improves PET/CT lesion segmentation (Dice 0.582 vs 0.543, Table 2). This is an empirical ablation: the same BiXLSTM–ResNet architecture is trained either from scratch or after MAE pretraining, and both are evaluated on the same AutoPET Task 1 benchmark. No equation or construction makes the SSL result equal to its input. The MAE method is introduced via citation to external literature ([13] AMAES, and the general MAE line) as well as the authors' own prior work ([3], [9]); however, these self-citations are not load-bearing in the sense of providing a uniqueness theorem or an ansatz that by itself forces the reported Dice values. The from-scratch baseline is not derived from the SSL model, nor is the reported improvement a renamed fitting parameter. The main weakness is statistical: the paper calls the improvement 'significant' without reporting standard deviations or significance tests, and the two SSL variants differ by only 0.002 Dice. That is a correctness/evidence concern, not a circularity concern, because the comparison remains an externally anchored, falsifiable experiment. There is no self-definitional step, no fitted input renamed as a prediction, and no imported uniqueness claim. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard, previously published components (MAE, XLSTM, nnU-Net) and on unverified modeling assumptions: that MAE pretraining transfers, that the unlabeled Task 2 CT data helps, and that the reported mean metrics are stable. No new physical or architectural entity is postulated. The main free parameters are the SSL variant choice and the unreported MAE and loss hyperparameters.

free parameters (4)

SSL initialization point = encoder versus bottleneck variants
Two pretrained variants differ only in which part of the network inherits MAE weights; the choice is made by hand in Section 3, Table 2.
MAE pretraining recipe (mask ratio, patch size, epochs, data mix) = not reported
Stage 1 is described in one sentence in Section 2.4; none of the MAE hyperparameters are given, yet the downstream gain is attributed to this stage.
Loss weighting for Dice plus cross-entropy = not reported
Section 2.4 says the loss is a combination but does not specify weights.
Patch size = 128x192x160
Taken from the nnUNet v2 framework defaults (Section 2.3, Table 1), not optimized in this paper.

assumptions (4)

domain assumption MAE self-supervised representations transfer to the downstream lesion segmentation task
The entire Stage 1 to Stage 2 pipeline in Section 2.4 rests on this transfer being beneficial; no failure analysis is provided.
domain assumption The AutoPET Task 1 metric set (mean Dice, FN volume, FP volume) is a stable and sufficient evaluation of segmentation quality
Section 3 reports only means over the 5 folds with no variance, so the evaluation rests on an assumption of stability.
domain assumption Unlabeled longitudinal CT scans from Task 2 improve Task 1 generalization
Section 2.2 includes Task 2 data in pretraining, but no ablation isolates its contribution.
domain assumption nnUNet v2 preprocessing and augmentation defaults are appropriate for whole-body PET/CT
Section 2.3 adopts the framework defaults without validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Masked Autoencoder Pretraining and BiXLSTM ResNet Architecture for PET/CT Tumor Segmentation." pith.science (2026). https://pith.science/paper/Z4TPHJDW

@misc{pith2026250902602,
  author       = {Pith},
  title        = {Pith review of: Masked Autoencoder Pretraining and BiXLSTM ResNet Architecture for PET/CT Tumor Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z4TPHJDW}},
  note         = {Machine review of arXiv:2509.02602}
}

read the original abstract

The accurate segmentation of lesions in whole-body PET/CT imaging is es-sential for tumor characterization, treatment planning, and response assess-ment, yet current manual workflows are labor-intensive and prone to inter-observer variability. Automated deep learning methods have shown promise but often remain limited by modality specificity, isolated time points, or in-sufficient integration of expert knowledge. To address these challenges, we present a two-stage lesion segmentation framework developed for the fourth AutoPET Challenge. In the first stage, a Masked Autoencoder (MAE) is em-ployed for self-supervised pretraining on unlabeled PET/CT and longitudinal CT scans, enabling the extraction of robust modality-specific representations without manual annotations. In the second stage, the pretrained encoder is fine-tuned with a bidirectional XLSTM architecture augmented with ResNet blocks and a convolutional decoder. By jointly leveraging anatomical (CT) and functional (PET) information as complementary input channels, the model achieves improved temporal and spatial feature integration. Evalua-tion on the AutoPET Task 1 dataset demonstrates that self-supervised pre-training significantly enhances segmentation accuracy, achieving a Dice score of 0.582 compared to 0.543 without pretraining. These findings high-light the potential of combining self-supervised learning with multimodal fu-sion for robust and generalizable PET/CT lesion segmentation. Code will be available at https://github.com/RespectKnowledge/AutoPet_2025_BxLSTM_UNET_Segmentation

Figures

Figures reproduced from arXiv: 2509.02602 by the authors.

**Figure 1.** Overview of the two-stage PET/CT lesion segmentation pipeline. Stage 1 employs a Masked Autoencoder (MAE) for self-supervised pretraining on unlabeled PET/CT and longitudinal CT scans to learn robust modality-specific features. Stage 2 fine-tunes the pretrained encoder with a BiXLSTM–ResNet architecture and a convolutional decoder on annotated PET/CT data, fusing PET and CT channels for voxelwise lesion segmentati… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 12 canonical work pages

[1]

‘Automated Lesion Segmentation in Whole-body PET/CT and Longitudinal (autopet/ct IV)’

Küstner, Thomas, Sergios Gatidis, Ornela Megne, Michael Ingrisch, Matthias Fab- ritius, Jakob Dexl, Katharina Jeblick, et al. ‘Automated Lesion Segmentation in Whole-body PET/CT and Longitudinal (autopet/ct IV)’. Zenodo, 18 March 2025. https://doi.org/10.5281/zenodo.15045096

work page doi:10.5281/zenodo.15045096 2025
[2]

Results from the autoPET challenge on fully automated lesion segmentation in oncologic PET/CT imaging

Gatidis, Sergios, Marcel Früh, Matthias P. Fabritius, Sijing Gu, Konstantin Niko- laou, Christian La Fougère, Jin Ye et al. "Results from the autoPET challenge on fully automated lesion segmentation in oncologic PET/CT imaging." Nature Ma- chine Intelligence 6, no. 11 (2024): 1396-1405

work page 2024
[3]

Mazher, M., Razzak, I., Qayyum, A., Tanveer, M., Beier, S., Khan, T., & Niederer, S. A. (2024). Self-supervised spatial–temporal transformer fusion based federated framework for 4D cardiovascular image segmentation. Information Fusion, 106, 102256

work page 2024
[4]

Extreme Cardiac MRI Analysis under Respiratory Motion: Results of the CMRxMotion Challenge

Wang, Kang, Chen Qin, Zhang Shi, Haoran Wang, Xiwen Zhang, Chen Chen, Cheng Ouyang et al. "Extreme Cardiac MRI Analysis under Respiratory Motion: Results of the CMRxMotion Challenge." arXiv preprint arXiv:2507.19165 (2025)

work page arXiv 2025
[5]

Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra

Yang, Kaiyuan, Fabio Musio, Yihui Ma, Norman Juchler, Johannes C. Paetzold, Rami Al-Maskari, Luciano Höher et al. "Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra." ArXiv (2024): arXiv-2312

work page 2024
[6]

ISLES'24: Final Infarct Prediction with Multimodal Imaging and Clinical Data. Where Do We Stand?

de la Rosa, Ezequiel, Ruisheng Su, Mauricio Reyes, Roland Wiest, Evamaria O. Riedel, Florian Kofler, Kaiyuan Yang et al. "Isles’ 24: improving final infarct pre- diction in ischemic stroke using multimodal imaging and clinical data (2024)." arXiv preprint arXiv:2408.10966

work page Pith review arXiv 2024
[7]

Multi -center fetal brain tissue annota- tion (feta) challenge 2022 results

Payette, Kelly, Céline Steger, Roxane Licandro, Priscille De Dumast, Hongwei Bran Li, Matthew Barkovich, Liu Li et al. "Multi -center fetal brain tissue annota- tion (feta) challenge 2022 results." IEEE transactions on medical imaging (2024)

work page 2022
[8]

Multi-class segmen- tation of aortic branches and zones in computed tomography angiography: The aortaseg24 challenge

Imran, Muhammad, Jonathan R. Krebs, Vishal Balaji Sivaraman, Teng Zhang, Amarjeet Kumar, Walker R. Ueland, Michael J. Fassler et al. "Multi-class segmen- tation of aortic branches and zones in computed tomography angiography: The aortaseg24 challenge." arXiv preprint arXiv:2502.05330 (2025)

arXiv 2025

Show all 14 references

[10]

Transforming 10 Heart Chamber Imaging: Self -Supervised Learning for Whole Heart Reconstruc- tion and Segmentation

Qayyum, Abdul, Hao Xu, Brian P. Halliday, Cristobal Rodero, Christopher W. Lanyon, Richard D. Wilkinson, and Steven Alexander Niederer. "Transforming 10 Heart Chamber Imaging: Self -Supervised Learning for Whole Heart Reconstruc- tion and Segmentation." arXiv preprint arXiv:24...

2024 arXiv
[11]

Hunting imaging biomarkers in pulmonary fibrosis: benchmarks of the AIIB23 challenge

Nan, Yang, Xiaodan Xing, Shiyi Wang, Zeyu Tang, Federico N. Felder, Sheng Zhang, Roberta Eufrasia Ledda et al. "Hunting imaging biomarkers in pulmonary fibrosis: benchmarks of the AIIB23 challenge." Medical Image Analysis 97 (2024): 103253

2024
[12]

A robust ensemble algorithm for ischemic stroke lesion segmentation: Generalizability and clinical utility beyond the isles challenge

de la Rosa, Ezequiel, Mauricio Reyes, Sook -Lei Liew, Alexandre Hutton, Roland Wiest, Johannes Kaesmacher, Uta Hanning et al. "A robust ensemble algorithm for ischemic stroke lesion segmentation: Generalizability and clinical utility beyond the isles challenge." arXiv preprint...

2024 arXiv
[13]

Amaes: Augmented masked autoencoder pretraining on public brain mri data for 3d-native segmentation

Munk, Asbjørn, Jakob Ambsdorf, Sebastian Llambias, and Mads Nielsen. "Amaes: Augmented masked autoencoder pretraining on public brain mri data for 3d-native segmentation." arXiv preprint arXiv:2408.00640 (2024)

2024 arXiv
[14]

nnU-Net: a self-configuring method for deep learning-based biomed- ical image segmentation

Isensee, Fabian, Paul F. Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H. Maier-Hein. "nnU-Net: a self-configuring method for deep learning-based biomed- ical image segmentation." Nature methods 18, no. 2 (2021): 203-211

2021
[15]

xlstm: Extended long short-term memory

Beck, Maximilian, Korbinian Pöppel, Markus Spanring, Andreas Auer, Oleksan- dra Prudnikova, Michael Kopp, Günter Klambauer, Johannes Brandstetter, and Sepp Hochreiter. "xlstm: Extended long short-term memory." Advances in Neural Information Processing Systems 37 (2024): 107547-107603

2024

Pith tools

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

[1] [1]

‘Automated Lesion Segmentation in Whole-body PET/CT and Longitudinal (autopet/ct IV)’

Küstner, Thomas, Sergios Gatidis, Ornela Megne, Michael Ingrisch, Matthias Fab- ritius, Jakob Dexl, Katharina Jeblick, et al. ‘Automated Lesion Segmentation in Whole-body PET/CT and Longitudinal (autopet/ct IV)’. Zenodo, 18 March 2025. https://doi.org/10.5281/zenodo.15045096

work page doi:10.5281/zenodo.15045096 2025

[2] [2]

Results from the autoPET challenge on fully automated lesion segmentation in oncologic PET/CT imaging

Gatidis, Sergios, Marcel Früh, Matthias P. Fabritius, Sijing Gu, Konstantin Niko- laou, Christian La Fougère, Jin Ye et al. "Results from the autoPET challenge on fully automated lesion segmentation in oncologic PET/CT imaging." Nature Ma- chine Intelligence 6, no. 11 (2024): 1396-1405

work page 2024

[3] [3]

Mazher, M., Razzak, I., Qayyum, A., Tanveer, M., Beier, S., Khan, T., & Niederer, S. A. (2024). Self-supervised spatial–temporal transformer fusion based federated framework for 4D cardiovascular image segmentation. Information Fusion, 106, 102256

work page 2024

[4] [4]

Extreme Cardiac MRI Analysis under Respiratory Motion: Results of the CMRxMotion Challenge

Wang, Kang, Chen Qin, Zhang Shi, Haoran Wang, Xiwen Zhang, Chen Chen, Cheng Ouyang et al. "Extreme Cardiac MRI Analysis under Respiratory Motion: Results of the CMRxMotion Challenge." arXiv preprint arXiv:2507.19165 (2025)

work page arXiv 2025

[5] [5]

Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra

Yang, Kaiyuan, Fabio Musio, Yihui Ma, Norman Juchler, Johannes C. Paetzold, Rami Al-Maskari, Luciano Höher et al. "Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra." ArXiv (2024): arXiv-2312

work page 2024

[6] [6]

ISLES'24: Final Infarct Prediction with Multimodal Imaging and Clinical Data. Where Do We Stand?

de la Rosa, Ezequiel, Ruisheng Su, Mauricio Reyes, Roland Wiest, Evamaria O. Riedel, Florian Kofler, Kaiyuan Yang et al. "Isles’ 24: improving final infarct pre- diction in ischemic stroke using multimodal imaging and clinical data (2024)." arXiv preprint arXiv:2408.10966

work page Pith review arXiv 2024

[7] [7]

Multi -center fetal brain tissue annota- tion (feta) challenge 2022 results

Payette, Kelly, Céline Steger, Roxane Licandro, Priscille De Dumast, Hongwei Bran Li, Matthew Barkovich, Liu Li et al. "Multi -center fetal brain tissue annota- tion (feta) challenge 2022 results." IEEE transactions on medical imaging (2024)

work page 2022

[8] [8]

Multi-class segmen- tation of aortic branches and zones in computed tomography angiography: The aortaseg24 challenge

Imran, Muhammad, Jonathan R. Krebs, Vishal Balaji Sivaraman, Teng Zhang, Amarjeet Kumar, Walker R. Ueland, Michael J. Fassler et al. "Multi-class segmen- tation of aortic branches and zones in computed tomography angiography: The aortaseg24 challenge." arXiv preprint arXiv:2502.05330 (2025)

arXiv 2025