Is Long Range Sequential Modeling Necessary For Colorectal Tumor Segmentation?

Abhishek Srivastava; Gorkem Durak; Gulsah Ozden; Koushik Biswas; Mustafa Adli; Ulas Bagci

REVIEW 3 major objections 5 minor 27 references

Is Long Range Sequential Modeling Necessary For Colorectal Tumor Segmentation?

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read For small colorectal tumors, local-only token modeling beats long-range sequence models.

desk verdict Useful new dataset and a plausible question, but the headline comparison is under-powered and the paper's own numbers don't line up. read the letter →

arxiv 2502.07120 v1 pith:JEQXELHA submitted 2025-02-10 cs.CV

Abhishek Srivastava , Koushik Biswas , Gorkem Durak , Gulsah Ozden , Mustafa Adli , Ulas Bagci This is my paper · ORCID

classification cs.CV

keywords colorectaltumorsegmentation3DCTMambastatespacemodelsTransformerlocaltokeninteractionCTS-204dataset

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 asks whether the long-range sequential modeling built into Mamba and Transformer architectures is actually necessary for tumor segmentation. The authors introduce CTS-204, a 204-patient CT dataset of colorectal tumors annotated by consensus, and propose MambaOutUNet, a UNet variant that removes the state-space token mixer and relies on gated convolutional local token interactions. Their central finding is that this local-only model beats all tested global-modeling baselines on CTS-204, while the same global models remain competitive on a multi-organ dataset where organs are larger and more variable in size. If the claim holds, it reorients 3D tumor segmentation research toward efficient local token modeling for small, complex regions of interest.

What carries the argument

The load-bearing component is the MambaOut block, a token-mixer substitution: where a Mamba block mixes tokens with $\mathrm{SSM}(\sigma(\mathrm{Conv}(x)))$, MambaOutUNet uses $\mathrm{Conv}(x)$ alone, dropping the sequential state-space model while keeping the layer norm and MLP around it. Placed in a SegMamba-style encoder-decoder with gated spatial convolution and feature-level uncertainty estimation, this one swap is what lets the comparison attribute any accuracy gap to long-range sequential modeling.

What would settle it

A direct falsifier would be a small-ROI tumor dataset on which a long-range model, tuned per architecture to convergence, matches or exceeds MambaOutUNet's Dice score; in particular, one could retrain SegMamba on CTS-204 with its own optimal schedule and check whether its 0.3990 score rises toward or past 0.5203.

Watch

Extended reading notes

Core claim

The paper claims that for segmenting colorectal tumors in 3D CT—targets that are small, irregular, and anatomically isolated—the long-range sequential modeling that Mamba and Transformer architectures are built for is not just unnecessary but currently a liability. On the new CTS-204 dataset, MambaOutUNet, which substitutes a gated convolution for the state-space token mixer while keeping the rest of the SegMamba-style UNet skeleton intact, reaches a Dice score of 0.5203, ahead of every tested baseline, including SegMamba (0.3990) and SwinUNet (0.4602). The authors read the BTCV multi-organ results, where Mamba- and Transformer-based models remain competitive, as evidence that global token modeling pays off only when the region of interest varies widely in size, and that local token interaction should be the default for small-ROI tumor segmentation.

Load-bearing premise

The ranking depends on the assumption that one shared training schedule—200 epochs and a single learning rate for every model—lets each architecture show its near-best performance; if SegMamba, SwinUNet, or nnU-Net are undertuned under that schedule, the conclusion about long-range modeling would not follow.

Editorial extensions

If this is right

For small, anatomically complex tumors, a convolutional local token mixer can be the highest-performing choice, so future architectures should be evaluated against a local-only baseline before adding global modules.
The CTS-204 dataset provides a benchmark where current long-range segmenters underperform, making it a useful stress test for claims about Mamba and Transformer gains in 3D segmentation.
On datasets with large and variable regions of interest, long-range modeling still appears desirable, so the two regimes likely call for different architectural defaults.
MambaOutUNet is proposed as a cheap, simple baseline: matching it is necessary before attributing gains to long-range sequence modeling.

Reading between the lines

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

An implicit corollary is that the optimal architecture depends on ROI size distribution: the paper's own BTCV results suggest a crossover point below which global context is wastage; measuring that crossover across datasets with graded tumor sizes would turn the hypothesis into a quantitative design rule.
The same local-versus-global trade-off likely applies to other small-target segmentation tasks such as brain metastases, early lung nodules, or pancreatic lesions, and one testable extension is to run the identical MambaOutUNet-versus-SegMamba comparison on those datasets.
If the claimed gap is real, then the recent trend of importing Mamba and SSM blocks into medical segmentation for small lesions may be adding compute without accuracy; a simpler check is whether removing the SSM from any Mamba-based segmenter preserves or improves its Dice score on a small-ROI dataset.

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

Dear colleague,

This paper asks the right question—whether long-range sequence modeling is actually needed for small-ROI tumor segmentation—and contributes a new colorectal tumor CT dataset, CTS-204, with 204 cases annotated by two radiation oncologists. That dataset alone makes the paper worth a look. Also to its credit, it puts a simple Gated-CNN UNet head-to-head against SegMamba, SwinUNet, and others on this data. The finding that MambaOutUNet wins on CTS-204 is genuinely new as a result, even though the architecture is a 3D repackaging of MambaOut's GatedCNN block.

Where it gets soft: the evidence doesn't yet support the strong conclusion. The numeric mismatch alone is enough to raise flags—Section 5 reports MambaOutUNet's mIoU and NSD on CTS-204 as 0.6971 and 0.6971, while Table 1 lists 0.3950 and 0.6139. That kind of inconsistency suggests careless error checking at best and sloppy reporting at worst.

The bigger issue is the fairness of the comparison. The paper states all methods were trained for 200 epochs with the same learning rate and validation interval. That's a strong assumption. nnU-Net is specifically designed to self-configure per dataset; forcing it into a shared schedule is known to handicap it. The fact that nnU-Net scores below plain UNet on CTS-204 (0.4842 vs. 0.5007) is a concrete red flag. Without error bars or repeated runs on the 21-case test split, the margin between MambaOutUNet (0.5203) and UNet (0.5007) is about two DSC points—inside the range of noise.

There's also a minor inconsistency between the promise of public release and the fine print "will be made available upon request," which isn't the same thing. And the BTCV results are described as supporting the long-range-is-desirable view, even though the local model actually wins there too. The paper would be stronger if it simply said local models are competitive across the board and then discussed when long-range helps.

Net: the dataset and the question deserve attention, but the central claim is not yet established. This is a conditional acceptance with major revisions, not a reject. The authors need to fix the reporting errors, redo baselines with their own tuned protocols or justify the shared schedule, and run repeated seeds with error bars.

I'd bring this to a reading group once the dataset is actually available and the numbers are cleaned up.

Referee Report

3 major / 5 minor

Summary. The paper introduces CTS-204, a new colorectal tumor CT segmentation dataset with 204 patient cases, and compares several long-range sequential modeling architectures (Mamba-based, Transformer-based, and hybrid variants) against a proposed local-gating architecture called MambaOutUNet. The authors report that MambaOutUNet achieves the best Dice score on CTS-204, and on the external BTCV multi-organ dataset it is also the best. On this basis they argue that for small, anatomically complex tumor ROIs, local token interactions are more useful than long-range sequence modeling, and they propose MambaOutUNet as a baseline for future volumetric tumor segmentation research.

Significance. If the comparison is fair and the reported numbers are correct, the paper makes a useful empirical contribution: it provides a new public dataset for colorectal tumor segmentation, and it challenges the common assumption that long-range modeling is automatically beneficial in 3D medical segmentation. The inclusion of an external dataset (BTCV) is a strength, as is the direct architectural comparison between Mamba, Transformer, and gated-convolution designs. However, the credibility of the central claim depends on the fairness of the training protocol and on internal consistency of the reported metrics, both of which currently have serious issues. The paper does not currently provide enough evidence to support a strong conclusion about the necessity or non-necessity of long-range sequential modeling.

major comments (3)

[Section 5, Table 1] The Results paragraph in Section 5 states that MambaOutUNet achieves mIoU and NSD of 0.6971 and 0.6971 on CTS-204, but Table 1 lists the same metrics as 0.3950 and 0.6139. The claimed margins over the second-best method (1.96%, 1.11%, 0.23%) are consequently not reproducible from the table. This is a load-bearing numerical inconsistency that must be resolved before the paper's performance claims can be assessed.
[Section 4, Table 1] The central comparison rests on training every method for 200 epochs with the same learning rate and validation interval, with no per-method tuning or hyperparameter search. The reported nnU-Net DSC of 0.4842, below plain UNet's 0.5007, is a concrete warning that a fixed shared protocol may be unfavorable to self-configuring or Mamba-based baselines. Since the margin between MambaOutUNet (0.5203) and the next best method (UNet, 0.5007) is only about two DSC points, and no error bars or significance tests are reported on a single 21-case test split, the ranking in Table 1 may reflect training choices rather than the intrinsic value of long-range modeling. The authors need to show that each baseline is at or near its own optimal setting, or at least to report sensitivity to the shared schedule.
[Section 5, Table 1] The discussion asserts that on BTCV, where the region of interest has high variance in size and structure, long-range sequencing is desirable, and that Mamba/Transformer models are competitive there. However, MambaOutUNet has the best DSC (0.8338), mIoU (0.7345), and NSD (0.9397) on BTCV as well. This undermines the proposed dichotomy between small-ROI and large-ROI tasks and requires a revised interpretation: either the local-gating mechanism is also beneficial on the large-ROI dataset, or the comparison on BTCV is not actually supportive of the paper's central hypothesis as stated.

minor comments (5)

[Section 2] The dataset description states that CTS-204 "will be made available upon request" but gives no access mechanism, license, or detailed preprocessing information such as CT windowing, resampling, or intensity normalization; a formal datasheet would improve reproducibility.
[Section 5] The sentence beginning "From Table 1 and Figure 1, we can observe..." uses "From" with a capital letter after a comma; please fix the capitalization and sentence flow.
[Section 1 and throughout] The name "MambaOutUNet" is written inconsistently as "MambaOutUnet", "MambaOutUNet", and "MambaOutUnet" at different points; please standardize the notation.
[Section 3.3, Eq. (8)] The definitions of "shift" and "flip" in Eq. (8) are informal; please state explicitly along which dimension the shift/zero-padding is applied and how the sequence is constructed from the 3D volume.
[References] Reference [15] is formatted incorrectly as "et al. Maier-Hein, Lena"; it should follow the standard author list and title format used by the journal.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the paper is an empirical architecture comparison with no fitted-parameter or self-citation load-bearing steps.

full rationale

This paper makes an empirical claim: on the new CTS-204 dataset, a Gated-CNN/UNet variant without SSM token mixing (MambaOutUNet) outperforms Transformer/Mamba baselines, and on BTCV long-range methods remain competitive. The numbers in Table 1 are produced by training fixed architectures under a shared protocol; no quantity used in the comparison is defined in terms of the reported outcome, and no fitted constant is later relabeled as a prediction. The design choice to ablate the SSM follows the protocol in Yu and Wang (MambaOut, [10]), which is an external prior work by different authors, not a self-citation. The central conclusion is not forced by any equation: there is no mathematical derivation from first principles, and the paper's only predictive content is the observed ranking, which is in principle falsifiable by rerunning the experiments. The strongest concern is experimental fairness (same 200-epoch schedule for all baselines, no error bars, and a two-point DSC margin), but that is a validity threat, not circularity. Also, Section 5 reports mIoU/NSD values (0.6971, 0.6971) that disagree with Table 1 (0.3950, 0.6139); this is an internal inconsistency that should be corrected, but it does not constitute a circular derivation. No self-citation chain, no uniqueness theorem, and no ansatz smuggled via citation is load-bearing.

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

This is an empirical comparison, so it introduces no fitted constants and no theoretical entities. The main burden is on fairness of the training and evaluation protocol, the accuracy of the new annotations, and the choice of metrics; these are captured as domain assumptions rather than free parameters.

assumptions (4)

domain assumption Equal training epochs and learning rate across all compared models yields a fair, representative comparison.
Invoked in Section 4 ('trained ... for 200 epochs for the same learning rate and validation interval') and used in Section 5. If baselines are undertuned, the performance ordering is an artifact of training schedule rather than architectural merit.
domain assumption The two radiation oncologists' consensus annotations in CTS-204 are an accurate ground truth for colorectal tumors.
Section 2 states annotations were made in consensus; no inter-observer variability or validation against pathology is reported.
domain assumption Dice, mIoU, and NSD computed following reference [15] adequately capture segmentation quality for the clinical conclusion.
Used in Section 5; no confidence intervals or statistical tests are provided to support the ordering.
domain assumption The Gated CNN block without SSM is a valid representative of local token interaction, isolating the effect of long-range modeling.
Section 3.5 defines TokenMixerGatedCNN = Conv(x) as the MambaOut ablation; this assumes all performance differences are due to the SSM token mixer alone rather than other implementation details.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Is Long Range Sequential Modeling Necessary For Colorectal Tumor Segmentation?." pith.science (2026). https://pith.science/paper/JEQXELHA

@misc{pith2026250207120,
  author       = {Pith},
  title        = {Pith review of: Is Long Range Sequential Modeling Necessary For Colorectal Tumor Segmentation?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEQXELHA}},
  note         = {Machine review of arXiv:2502.07120}
}

read the original abstract

Segmentation of colorectal cancer (CRC) tumors in 3D medical imaging is both complex and clinically critical, providing vital support for effective radiation therapy planning and survival outcome assessment. Recently, 3D volumetric segmentation architectures incorporating long-range sequence modeling mechanisms, such as Transformers and Mamba, have gained attention for their capacity to achieve high accuracy in 3D medical image segmentation. In this work, we evaluate the effectiveness of these global token modeling techniques by pitting them against our proposed MambaOutUNet within the context of our newly introduced colorectal tumor segmentation dataset (CTS-204). Our findings suggest that robust local token interactions can outperform long-range modeling techniques in cases where the region of interest is small and anatomically complex, proposing a potential shift in 3D tumor segmentation research.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 18 canonical work pages

[1]

However, this task presents several challenges, variability in image quality and the complex anatomy of the colorectal region hinders precise delineation of tumors

INTRODUCTION Colorectal cancer (CRC) tumor segmentation from medical scans, such as computed tomography (CT), is a crucial task in the field of medical imaging and computer-assisted diagno- sis. However, this task presents several challenges, variability in image quality and the complex anatomy of the colorectal region hinders precise delineation of tumor...

work page
[2]

Is Long-Range Sequence Modeling Necessary for Tu- mor Segmentation? While image segmentation is essen- tially a long range sequence modeling task, in case of tumor segmentation where the target region by its natural properties has no correlation with most of the surrounding pixels/voxels, does it still holds the same efficacy? We investigate the po- tenti...

work page
[3]

We propose a newly curated CTS-204 dataset for col- orectal cancer tumor segmentation

work page
[4]

Our results demonstrate that efficient channel mixing and spatially gated features can outperform many existing computationally intensive long- range modeling techniques

Comprehensive Analysis of Long Range Sequence Modeling Architectures: We propose a comprehensive analysis of recent Mamba-based architectures, including UNet variations that enhance the capacity to model local- and global token interactions. Our results demonstrate that efficient channel mixing and spatially gated features can outperform many existing com...

work page
[5]

Is Long Range Sequential Modeling Necessary For Colorectal Tumor Segmentation?

DA TASET We have publicly released a new colorectal tumor segmen- tation dataset termed ”CTS-204”. The dataset was collected arXiv:2502.07120v1 [cs.CV] 10 Feb 2025 Table 1: Quantitative Comparison on the CTS 204 and Synapse multi-organ CT dataset (BTCV). We report the Dice Score (DSC), Mean Intersection over Union (mIoU), and Normalized Surface Distance (...

work page Pith review arXiv 2025
[6]

Assessing the impact of long range sequence modeling through Mamba requires a fair comparison with SegMamba [9]

METHODS In this section, we describe the architecture of our three pro- posed baselines: Mamba-SwinUnet, SegHydra, and Mam- baOutUnet. Assessing the impact of long range sequence modeling through Mamba requires a fair comparison with SegMamba [9]. We ensure this by following the same archi- tecture as proposed in SegMamba [9] while invoking neces- sary ar...

work page
[7]

The relevant baselines and all proposed architectures were trained on CTS- 204 for 200 epochs for the same learning rate and valida- tion interval

EXPERIMENTS CTS-204 was split into 3 splits, training, validation and test- ing which had 163, 20, 21 cases, respectively. The relevant baselines and all proposed architectures were trained on CTS- 204 for 200 epochs for the same learning rate and valida- tion interval. Similarly, all methodologies were trained on the BTCV dataset for 400 epochs on a sing...

work page
[8]

The results of these experiments are summarized in Tables 1

RESULTS AND DISCUSSION To validate our hypotheses, we conducted experiments on two datasets: the CTS 204 dataset and the BTCV multi-organ segmentation dataset. The results of these experiments are summarized in Tables 1. Here we can observe that the Mam- baOutUNet achieves a (DSC), mIoU, and NSD of 0.5203, 0.6971, and 0.6971 outperforming the second best ...

work page

Show all 27 references

[9]

CONCLUSION In this paper, we release CTS-204, a new colorectal tumor segmentation dataset with 204 distinct cases. We demonstrate the effectiveness of our MambaOut architecture for colorec- tal tumor segmentation which provides critical insights and a robust framework for futu...
[10]

Approval was granted by Northwest- ern University (No

COMPLIANCE WITH ETHICAL STANDARDS This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by Northwest- ern University (No. STU00214545)
[11]

ACKNOWLEDGMENTS This project is supported by NIH funding: R01-CA246704, R01-CA240639, U01-DK127384-02S1, and U01-CA268808
[12]

U-Net: convolutional networks for biomedical image segmentation,

Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-Net: convolutional networks for biomedical image segmentation,” in Proc. of Internat. Confer . on Med. Ima. Compu. Comput.-Assis. Interven. , 2015, pp. 234– 241

2015
[13]

Swin- unet: Unet-like pure transformer for medical image seg- mentation,

Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xi- aopeng Zhang, Qi Tian, and Manning Wang, “Swin- unet: Unet-like pure transformer for medical image seg- mentation,” in European conference on computer vision. Springer, 2022, pp. 205–218

2022
[14]

Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,

Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu, “Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,” in International MICCAI brain- lesion workshop. Springer, 2021, pp. 272–284

2021
[15]

An image is worth 16x16 words: Transformers for image recognition at scale,

Alexey Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929, 2020

2010 arXiv
[16]

Attention is all you need,

A Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

2017
[17]

Mamba: Linear-time sequence modeling with selective state spaces,

Albert Gu and Tri Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752, 2023

2023 arXiv
[18]

U-mamba: Enhancing long-range dependency for biomedical image segmenta- tion,

Jun Ma, Feifei Li, and Bo Wang, “U-mamba: Enhancing long-range dependency for biomedical image segmenta- tion,” arXiv preprint arXiv:2401.04722, 2024

2024 arXiv
[19]

nnu-net: a self- configuring method for deep learning-based biomedical image segmentation,

Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein, “nnu-net: a self- configuring method for deep learning-based biomedical image segmentation,” Nature methods, vol. 18, no. 2, pp. 203–211, 2021

2021
[20]

Segmamba: Long-range sequential model- ing mamba for 3d medical image segmentation,

Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu, “Segmamba: Long-range sequential model- ing mamba for 3d medical image segmentation,” in In- ternational Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 2024, pp. 578–588

2024
[21]

Mambaout: Do we really need mamba for vision?,

Weihao Yu and Xinchao Wang, “Mambaout: Do we really need mamba for vision?,” arXiv preprint arXiv:2405.07992, 2024

2024 arXiv
[22]

Uncertainty-aware multi-dimensional mutual learning for brain and brain tumor segmentation,

Junting Zhao, Zhaohu Xing, Zhihao Chen, Liang Wan, Tong Han, Huazhu Fu, and Lei Zhu, “Uncertainty-aware multi-dimensional mutual learning for brain and brain tumor segmentation,” IEEE Journal of Biomedical and Health Informatics, vol. 27, no. 9, pp. 4362–4372, 2023

2023
[23]

Transformers are ssms: Generalized models and efficient algorithms through structured state space duality,

Tri Dao and Albert Gu, “Transformers are ssms: Generalized models and efficient algorithms through structured state space duality,” arXiv preprint arXiv:2405.21060, 2024

2024 arXiv
[24]

Mambavision: A hybrid mamba-transformer vision backbone,

Ali Hatamizadeh and Jan Kautz, “Mambavision: A hybrid mamba-transformer vision backbone,” arXiv preprint arXiv:2407.08083, 2024

2024 arXiv
[25]

Language modeling with gated convolu- tional networks,

Yann N Dauphin, Angela Fan, Michael Auli, and David Grangier, “Language modeling with gated convolu- tional networks,” in International conference on ma- chine learning. PMLR, 2017, pp. 933–941

2017
[26]

Metrics reloaded: Pitfalls and recommendations for image analysis validation,

et al. Maier-Hein, Lena, “Metrics reloaded: Pitfalls and recommendations for image analysis validation,”Nature methods, 21(2), 195-212, 2024

2024
[27]

Hydra: Bidirectional state space models through generalized matrix mixers,

Sukjun Hwang, Aakash Lahoti, Tri Dao, and Al- bert Gu, “Hydra: Bidirectional state space models through generalized matrix mixers,” arXiv preprint arXiv:2407.09941, 2024

2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

However, this task presents several challenges, variability in image quality and the complex anatomy of the colorectal region hinders precise delineation of tumors

INTRODUCTION Colorectal cancer (CRC) tumor segmentation from medical scans, such as computed tomography (CT), is a crucial task in the field of medical imaging and computer-assisted diagno- sis. However, this task presents several challenges, variability in image quality and the complex anatomy of the colorectal region hinders precise delineation of tumor...

work page

[2] [2]

Is Long-Range Sequence Modeling Necessary for Tu- mor Segmentation? While image segmentation is essen- tially a long range sequence modeling task, in case of tumor segmentation where the target region by its natural properties has no correlation with most of the surrounding pixels/voxels, does it still holds the same efficacy? We investigate the po- tenti...

work page

[3] [3]

We propose a newly curated CTS-204 dataset for col- orectal cancer tumor segmentation

work page

[4] [4]

Our results demonstrate that efficient channel mixing and spatially gated features can outperform many existing computationally intensive long- range modeling techniques

Comprehensive Analysis of Long Range Sequence Modeling Architectures: We propose a comprehensive analysis of recent Mamba-based architectures, including UNet variations that enhance the capacity to model local- and global token interactions. Our results demonstrate that efficient channel mixing and spatially gated features can outperform many existing com...

work page

[5] [5]

Is Long Range Sequential Modeling Necessary For Colorectal Tumor Segmentation?

DA TASET We have publicly released a new colorectal tumor segmen- tation dataset termed ”CTS-204”. The dataset was collected arXiv:2502.07120v1 [cs.CV] 10 Feb 2025 Table 1: Quantitative Comparison on the CTS 204 and Synapse multi-organ CT dataset (BTCV). We report the Dice Score (DSC), Mean Intersection over Union (mIoU), and Normalized Surface Distance (...

work page Pith review arXiv 2025

[6] [6]

Assessing the impact of long range sequence modeling through Mamba requires a fair comparison with SegMamba [9]

METHODS In this section, we describe the architecture of our three pro- posed baselines: Mamba-SwinUnet, SegHydra, and Mam- baOutUnet. Assessing the impact of long range sequence modeling through Mamba requires a fair comparison with SegMamba [9]. We ensure this by following the same archi- tecture as proposed in SegMamba [9] while invoking neces- sary ar...

work page

[7] [7]

The relevant baselines and all proposed architectures were trained on CTS- 204 for 200 epochs for the same learning rate and valida- tion interval

EXPERIMENTS CTS-204 was split into 3 splits, training, validation and test- ing which had 163, 20, 21 cases, respectively. The relevant baselines and all proposed architectures were trained on CTS- 204 for 200 epochs for the same learning rate and valida- tion interval. Similarly, all methodologies were trained on the BTCV dataset for 400 epochs on a sing...

work page

[8] [8]

The results of these experiments are summarized in Tables 1

RESULTS AND DISCUSSION To validate our hypotheses, we conducted experiments on two datasets: the CTS 204 dataset and the BTCV multi-organ segmentation dataset. The results of these experiments are summarized in Tables 1. Here we can observe that the Mam- baOutUNet achieves a (DSC), mIoU, and NSD of 0.5203, 0.6971, and 0.6971 outperforming the second best ...

work page