Pith. sign in

REVIEW 3 major objections 6 minor 40 references

TransForSeg claims a stereo Vision Transformer can jointly segment catheters from two X-ray views and estimate 3D contact forces in a single end-to-end pass, outperforming the prior multitask CNN baseline on every benchmark dataset.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A shared-weight stereo ViT with cross-attention fusion segments the catheter in two views and regresses 3D tip forces, claiming state-of-the-art results on synthetic X-ray datasets.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection TransForSeg is a plausible first ViT multitask architecture for stereo catheter segmentation and 3D force estimation, but the SOTA claim rests on a synthetic-only benchmark with no variance estimates. the 3 major comments →

arxiv 2509.01605 v1 pith:2DLKQF5A submitted 2025-09-01 cs.CV cs.AIcs.LGcs.RO

TransForSeg: A Multitask Stereo ViT for Joint Stereo Segmentation and 3D Force Estimation in Catheterization

Pedram Fekri , Mehrdad Zadeh , Javad Dargahi This is my paper
classification cs.CV cs.AIcs.LGcs.RO
keywords stereo vision transformercatheter segmentation3D force estimationmultitask learningcross-attentionsynthetic X-raysemantic segmentationfluoroscopy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper proposes TransForSeg, a stereo Vision Transformer that takes two X-ray images of a catheter from different angles and, in one forward pass, outputs a segmentation map for each view plus a 3D force vector for the contact force at the catheter tip. The claim is that a patch-based transformer with a shared encoder-decoder and cross-attention between views captures the catheter's deflection better than the previous CNN-based multitask model, H-Net, without needing a separate segmentation step before force estimation. Across three datasets (RGB and two synthetic X-ray variants), TransForSeg reports lower force-estimation MSE than H-Net, with 51%, 38%, and 22% improvements respectively, and the highest segmentation mIoU and mDice scores on all three datasets. If correct, this makes sensor-free force estimation and catheter localization a single lightweight end-to-end network, which matters for real-time robotic catheterization. The paper also reports that the model retains competitive performance under several noise types not seen during training.

Core claim

On its own terms, the paper establishes that a stereo ViT can jointly solve two tasks previously handled by separate or CNN-based pipelines: pixel-level catheter segmentation from two viewpoints and regression of the 3D contact force at the tip. The architecture treats one view as encoder input and the other as decoder input, lets the decoder cross-attend to the encoder's tokens, and uses the resulting [CLS] token for force regression while sharing a CNN upsampler to produce both segmentation maps. The central quantitative claim is that this design outperforms H-Net, the prior multitask CNN, on all benchmark datasets in both tasks, and also beats single-task segmentation models such as DeepL

What carries the argument

The central object is the stereo Vision Transformer encoder-decoder with weight sharing and a cross-attention fusion block. Each X-ray view is cut into non-overlapping patches and embedded as a token sequence; the encoder processes one view, the weight-shared decoder processes the other, and multi-head cross-attention lets the decoder's tokens attend to the encoder's tokens so the model can register corresponding catheter deflections across the two angles. The [CLS] token of the decoder aggregates that cross-view information for the MLP force head, while patch tokens from both encoder and decoder are reshaped and upsampled by a shared CNN decoder into two segmentation maps. The load-bearing

Load-bearing premise

The synthetic X-ray images made from RGB photos mimic real fluoroscopy closely enough that scores on them predict how the model will behave in an actual catheterization room.

What would settle it

Collect real biplanar fluoroscopy of a catheter pressed against a force sensor and run the trained TransForSeg model; if its force MSE is an order of magnitude above the synthetic-X-ray test value, or if its segmentation masks lose the catheter in anatomy-heavy frames, the claimed state-of-the-art performance has not transferred to clinical conditions.

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

If this is right

  • A single end-to-end network can replace the two-stage pipeline of first segmenting the catheter and then estimating force from the cleaned image, simplifying both training and deployment.
  • Because the segmentation task provides auxiliary supervision, force estimation improves on synthetic X-ray data even though force regression alone is the harder-to-converge head.
  • Sharing weights between encoder and decoder and sharing the CNN upsampler keeps the model lightweight despite having two segmentation heads and a regression head.
  • The model maintains near-baseline segmentation and force performance under impulse, Poisson, and stripe noise on synthetic X-ray data, suggesting robustness to some real-world image corruptions.
  • The reported state-of-the-art numbers set a new comparison point for future multitask catheter perception models.

Where Pith is reading between the lines

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

  • The paper leaves untested whether the synthetic X-ray generator faithfully mimics real fluoroscopy; if it does, the same architecture should transfer to clinical data with little modification, but that transfer is a separate claim requiring a real-X-ray benchmark.
  • The same weight-shared cross-attention stereo design could be applied to other biplanar instrument-tracking problems, such as guidewire or endoscope tip force estimation, wherever a single regression target depends on the relative deflection between two views.
  • A testable extension is to use the segmentation head's multi-view consistency as a self-supervised signal on unlabeled real fluoroscopy, which could narrow the domain gap without requiring force-sensor ground truth.
  • The RGB ablation result implies that auxiliary segmentation is not universally beneficial: when background cues correlate with the target variable, segmentation can hurt force estimation, so multitask benefits should be verified per domain rather than assumed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes TransForSeg, a multitask stereo Vision Transformer for simultaneous catheter segmentation from two X-ray/RGB views and 3D contact-force estimation. Two images are processed as separate patch sequences, one through a ViT encoder and one through a weight-shared ViT decoder; cross-attention fuses the views, a CLS token feeds an MLP force head, and patch embeddings are reshaped and upsampled by a shared CNN decoder for segmentation. The model is trained end-to-end on RGB and two synthetic X-ray datasets derived from the authors' earlier Y-Net/H-Net framework. Experiments compare force estimation against MLP, ResNet, Y-Net, and H-Net, and segmentation against FCN, U-Net, HRNet, DeepLabV3, SegFormer, and MaskFormer. The paper reports force MSE improvements of 22-51% over H-Net and the highest mIoU/mDice across all three datasets, and includes an ablation (TransForcer) and a noise-robustness study.

Significance. If the results hold, TransForSeg is a meaningful architectural contribution: it is, to the authors' knowledge, the first multitask stereo ViT for joint catheter segmentation and 3D force estimation, and it demonstrates that weight sharing between encoder and decoder plus cross-attention fusion can be computationally efficient (Tiny: 6.9M params, 2.8 GFLOPs) while improving over CNN-based multitask baselines. The paper includes useful elements: three-run averaging, an ablation isolating the segmentation task, and a six-type perturbation study. These are strengths. However, the quantitative case is weakened by the absence of any uncertainty quantification and by the fact that all evaluation is on synthetic data from the authors' own generation pipeline; the clinical claim is therefore not yet established. The central benchmark claims are plausible, but currently supported only as synthetic-benchmark results.

major comments (3)
  1. [§IV.B, Tables I–II] All reported numbers are averages of three runs, but no standard deviations, confidence intervals, or significance tests are given. This is load-bearing for the claim that TransForSeg 'consistently outperforms' baselines. The segmentation task is near-saturated: mIoU values are around 98.4–98.7 for TransForSeg vs. 97.5–98.6 for SegFormer and 95.7 for H-Net. Differences of 0.1–0.3 mIoU may be within run-to-run variation, especially with only three runs. Please report per-run values and mean±std, and where possible a paired significance test or confidence intervals, for the main tables.
  2. [§IV datasets; §V conclusion] All experiments are performed on RGB images and on synthetic X-ray images generated by the H-Net pipeline, not on real fluoroscopy. Table IV's 'Domain Shift' experiments add synthetic corruptions to the same synthetic images; they do not model patient anatomy, X-ray scatter, quantum noise, acquisition geometry, or device artifacts. The conclusion explicitly defers real-world adaptation to future work. Consequently, the abstract and conclusion's 'new state-of-the-art in catheter segmentation and force estimation' is supported only for this synthetic benchmark, and the transferability to clinical catheterization rests on an untested faithfulness assumption. Please either scope the claims to synthetic benchmark performance or add a real/external validation dataset (or a much more realistic X-ray simulation).
  3. [§IV.C, Table III] The ablation's main conclusion is dataset-dependent. On RGB, TransForSeg without segmentation heads achieves substantially lower MSE than with them (Tiny: 7.75e-6 vs. 1.77e-5; Base: 7.09e-6 vs. 1.81e-5), i.e., the segmentation heads hurt force estimation by more than half. The improvements are mostly on the synthetic X-ray datasets (up to 28.6% for X-Ray1 Tiny and 20.66% for X-Ray2 Small). The paper does acknowledge this in the text, but the abstract and conclusion state that segmentation 'significantly enhances force prediction' and that the segmentation task 'guides the network to focus on the catheter’s deflection shape.' This claim needs to be explicitly scoped to the synthetic X-ray domains, or the proposed mechanism needs to be reconciled with the RGB result.
minor comments (6)
  1. [§III Patch Embeddings] The text says 'p × p patches' and 'total of +1 tokens' but should read 'p^2 patches' and 'total of p^2 + 1 tokens' (the notation p2+1 is used later; make it consistent).
  2. [§III Force Estimation Head vs. §IV.A] The force head is described as a '3-layer MLP with 64, 32, and 3 units' in Section III, but Section IV.A calls it a 'two-layer MLP' with projections to 64 and 32. Please reconcile the terminology.
  3. [Table IV, Defocus rows] The RMSE values for RGB Tiny and Base under Defocus are listed as 0.076 and 0.065, which appear to be an order of magnitude too large (likely 0.0076 and 0.0065). Also, the text has 'x2.1x' in one row, a typo.
  4. [§III, Eq. (3) paragraph] Minor typos: 'Each each attends' and 'where where Q = embs and k, V= embt' should be cleaned up.
  5. [§IV.C, Fig. 4] The text says 'Fig 4 illustrates the quantitative performance' but the figure shows qualitative segmentation outputs. Please correct the reference or the description.
  6. [General] The abstract and conclusion say 'new state-of-the-art in both catheter segmentation and force estimation' without qualification. Please add 'on the synthetic benchmark datasets used here' to keep the claims commensurate with the evaluation.

Circularity Check

0 steps flagged

No circular derivation found; mild self-referential benchmarking does not reduce the claims to their inputs.

full rationale

TransForSeg does not derive its outputs from its inputs by construction. Force targets come from a bench force sensor ('a standard catheter was pressed against a force sensor while two cameras captured its deflection from different angles'), and segmentation masks are part of the dataset; both enter only through the supervised losses in Eqs. (6) and (7). The reported MSE/mIoU numbers are measured on held-out test splits (13,650 training / 2,925 validation / 2,925 test samples), so they are empirical comparisons, not fitted parameters renamed as predictions. The architecture's design choices—patch embeddings, shared ViT encoder/decoder, cross-attention fusion, [CLS]-token MLP, and shared upsampler—are not defined in terms of the ground-truth force vector or segmentation masks. The strongest self-referential element is that the datasets and the principal multitask baseline (H-Net) come from the authors' prior work, and the force-estimation baselines (MLP [12], ResNet [14], Y-Net [6], H-Net [8]) are all from the same research group; this weakens the external benchmarking of the 'new state-of-the-art' force-estimation claim but is not a circular reduction. The paper itself flags the relevant limitation: 'Future work will explore its adaptation to real-world clinical settings and integration with autonomous robotic systems.' Synthetic-only evaluation is a generalizability and external-validity risk, not an equation-level loop. No uniqueness theorem or ansatz is imported from the authors' own citations to force the conclusion, and the comparison with external segmentation models (FCN, U-Net, HR-Net, DeepLabV3, SegFormer, MaskFormer) provides independent content. I therefore find no circular step that reduces a prediction to its inputs by definition.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The main empirical claims rest on the realism of the synthetic X-ray data and the validity of the force ground truth, both inherited from the authors' prior work (Y-Net, H-Net). Additional assumptions are the transferability of ImageNet pretraining and the benign effect of weight tying.

free parameters (1)
  • Loss weights λ1, λ2, λ3 = 1, 1, 1
    Set to equal weights in loss Equation (7); reported as 'empirically set to 1' without any tuning or sensitivity analysis (Section IV.A). The balance between force MSE and two segmentation BCE losses is fixed by hand.
axioms (4)
  • domain assumption The synthetic X-ray generator and experimental force measurements from Y-Net/H-Net accurately capture the physics of catheter deflection and contact force.
    All training and evaluation use datasets from [6], [8]; no independent clinical validation is offered. This is a premise of the entire empirical section (Section IV).
  • domain assumption ImageNet-pretrained ViT weights transfer sufficiently to synthetic X-ray images for both segmentation and force regression.
    Both model variants are initialized with ImageNet weights (Section IV.A) and fine-tuned; the paper does not compare against training from scratch or domain-specific pretraining.
  • ad hoc to paper Weight sharing between the encoder and decoder does not substantially degrade the model's ability to learn view-specific features.
    The ViT decoder shares weights with the encoder (Section III) for efficiency; no ablation tests weight tying vs. independent weights.
  • standard math Standard transformer mathematics (attention, layer normalization, residual connections) is valid for image patch sequences.
    Equations (1)-(5) rely on established attention and normalization formulations from [27] and [9].

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TransForSeg: A Multitask Stereo ViT for Joint Stereo Segmentation and 3D Force Estimation in Catheterization." pith.science (2026). https://pith.science/paper/2DLKQF5A

@misc{pith2026250901605,
  author       = {Pith},
  title        = {Pith review of: TransForSeg: A Multitask Stereo ViT for Joint Stereo Segmentation and 3D Force Estimation in Catheterization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DLKQF5A}},
  note         = {Machine review of arXiv:2509.01605}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recently, the emergence of multitask deep learning models has enhanced catheterization procedures by providing tactile and visual perception data through an end-to-end architecture. This information is derived from a segmentation and force estimation head, which localizes the catheter in X-ray images and estimates the applied pressure based on its deflection within the image. These stereo vision architectures incorporate a CNN-based encoder-decoder that captures the dependencies between X-ray images from two viewpoints, enabling simultaneous 3D force estimation and stereo segmentation of the catheter. With these tasks in mind, this work approaches the problem from a new perspective. We propose a novel encoder-decoder Vision Transformer model that processes two input X-ray images as separate sequences. Given sequences of X-ray patches from two perspectives, the transformer captures long-range dependencies without the need to gradually expand the receptive field for either image. The embeddings generated by both the encoder and decoder are fed into two shared segmentation heads, while a regression head employs the fused information from the decoder for 3D force estimation. The proposed model is a stereo Vision Transformer capable of simultaneously segmenting the catheter from two angles while estimating the generated forces at its tip in 3D. This model has undergone extensive experiments on synthetic X-ray images with various noise levels and has been compared against state-of-the-art pure segmentation models, vision-based catheter force estimation methods, and a multitask catheter segmentation and force estimation approach. It outperforms existing models, setting a new state-of-the-art in both catheter segmentation and force estimation.

Figures

Figures reproduced from arXiv: 2509.01605 by Javad Dargahi, Mehrdad Zadeh, Pedram Fekri.

Figure 1
Figure 1. Figure 1: The diagram shows TransForSeg’s architecture. It is equipped with a Vision Transformer (ViT)-based encoder and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The diagram shows the error histograms for the force estimation head of TransForSeg (’w’ for models with segmentation [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The diagram presents the training and validation loss curves. Subplots (a)–(d) display the segmentation loss for both [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results across 3 out of 6 noise conditions for 3 input samples. The first row displays the original (clean) [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 25 canonical work pages · 3 internal anchors

  1. [1]

    Y . R. Manda and K. M. Baradhi, Cardiac Catheterization Risks and Complications. StatPearls Publishing, Treasure Island (FL), 2021. [Online]. Available: http://europepmc.org/books/NBK531461

    work page 2021

  2. [2]

    Transradial cardiac catheterization: A review of access site complications,

    Y . Kanei, T. Kwan, N. C. Nakra, M. Liou, Y . Huang, L. L. Vales, J. T. Fox, J. P. Chen, and S. Saito, “Transradial cardiac catheterization: A review of access site complications,” Catheterization and Cardiovascular Interventions , vol. 78, no. 6, pp. 840–846, 2011. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/ccd. 22978

    work page doi:10.1002/ccd 2011

  3. [3]

    Toward autonomous cardiac catheterization through a parametric finite element simulation with experimental validation,

    M. Roshanfar, P. Fekri, and J. Dargahi, “Toward autonomous cardiac catheterization through a parametric finite element simulation with experimental validation,” ICAS 2023, p. 23, 3 2023

    work page 2023

  4. [4]

    Reinforcement learning for safe autonomous two-device navigation of cerebral vessels in mechanical thrombectomy,

    H. Robertshaw, B. Jackson, J. Wang, H. Sadati, L. Karstensen, A. Granados, and T. C. Booth, “Reinforcement learning for safe autonomous two-device navigation of cerebral vessels in mechanical thrombectomy,” International Journal of Computer Assisted Radiology and Surgery , Apr 2025. [Online]. Available: https: //doi.org/10.1007/s11548-025-03339-8

    doi:10.1007/s11548-025-03339-8 2025

  5. [5]

    Deep learning-based haptic guidance for surgical skills transfer,

    P. Fekri, J. Dargahi, and M. Zadeh, “Deep learning-based haptic guidance for surgical skills transfer,” Frontiers in Robotics and AI , vol. 7, 2021. [Online]. Available: https://www.frontiersin.org/article/10. 3389/frobt.2020.586707

    work page arXiv 2021

  6. [6]

    Y-net: A deep convolutional architecture for 3d estimation of contact forces in intracardiac catheters,

    P. Fekri, H. Khodashenas, K. Lachapelle, R. Cecere, M. Zadeh, and J. Dargahi, “Y-net: A deep convolutional architecture for 3d estimation of contact forces in intracardiac catheters,” IEEE Robotics and Automa- tion Letters, vol. 7, no. 2, pp. 3592–3599, 2022. 10

    work page 2022

  7. [7]

    End-to-End Real-time Catheter Segmentation with Optical Flow-Guided Warping during Endovascular Intervention

    A. Nguyen, D. Kundrat, G. Dagnino, W. Chi, M. E. M. K. Abdelaziz, Y . Guo, Y . Ma, T. M. Y . Kwok, C. Riga, and G.-Z. Yang, “End-to-end real-time catheter segmentation with optical flow-guided warping during endovascular intervention,” 2020. [Online]. Available: https://arxiv.org/abs/2006.09117

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  8. [8]

    H-net: A multitask architecture for simultaneous 3d force estimation and stereo semantic segmentation in intracardiac catheters,

    P. Fekri, M. Zadeh, and J. Dargahi, “H-net: A multitask architecture for simultaneous 3d force estimation and stereo semantic segmentation in intracardiac catheters,” IEEE Robotics and Automation Letters , vol. 10, no. 1, pp. 844–851, 2025

    work page 2025

  9. [9]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” CoRR, vol. abs/2010.11929, 2020. [Online]. Available: https://arxiv.org/abs/2010.11929

    Pith/arXiv arXiv 2010

  10. [10]

    Unknown Object Segmentation from Stereo Images

    M. Durner, W. Boerdijk, M. Sundermeyer, W. Friedl, Z. Marton, and R. Triebel, “Unknown object segmentation from stereo images,” CoRR, vol. abs/2103.06796, 2021. [Online]. Available: https://arxiv.org/abs/ 2103.06796

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  11. [11]

    Revisiting stereo depth estimation from a sequence- to-sequence perspective with transformers,

    Z. Li, X. Liu, N. Drenkow, A. Ding, F. X. Creighton, R. H. Taylor, and M. Unberath, “Revisiting stereo depth estimation from a sequence- to-sequence perspective with transformers,” in 2021 IEEE/CVF Interna- tional Conference on Computer Vision (ICCV) , 2021, pp. 6177–6186

    work page 2021

  12. [12]

    A vision-based method for estimating contact forces in intracardiac catheters,

    H. Khodashenas, P. Fekri, M. Zadeh, and J. Dargahi, “A vision-based method for estimating contact forces in intracardiac catheters,” in IEEE 1st International Conference on Autonomous Systems, ICAS . IEEE, 2021

    work page 2021

  13. [13]

    A deep learning model for tip force estimation on steerable catheters via learning-from-simulation,

    M. Roshanfar, P. Fekri, and J. Dargahi, “A deep learning model for tip force estimation on steerable catheters via learning-from-simulation,” EasyChair, no. 10364, 6 2023

    work page 2023

  14. [14]

    A deep learning force estimator system for intracardiac catheters,

    P. Fekri, H. R. Nourani, M. Razban, J. Dargahi, M. Zadeh, and A. Arshi, “A deep learning force estimator system for intracardiac catheters,” in 2021 IEEE International Symposium on Medical Measurements and Applications (MeMeA). IEEE, 2021, pp. 1–6

    work page 2021

  15. [15]

    Deep reinforcement learning for the navigation of neurovascular catheters,

    T. Behr, T. P. Pusch, M. Siegfarth, D. H ¨usener, T. M ¨orschel, and L. Karstensen, “Deep reinforcement learning for the navigation of neurovascular catheters,” Current Directions in Biomedical Engineering, vol. 5, no. 1, pp. 5–8, 2019. [Online]. Available: https://doi.org/10. 1515/cdbme-2019-0002

    work page 2019

  16. [16]

    A Novel Sample-efficient Deep Reinforcement Learning with Episodic Policy Transfer for PID-Based Control in Cardiac Catheterization Robots

    O. M. Omisore, T. Akinyemi, W. Duan, W. Du, and L. Wang, “A novel sample-efficient deep reinforcement learning with episodic policy transfer for pid-based control in cardiac catheterization robots,” 2021. [Online]. Available: https://arxiv.org/abs/2110.14941

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  17. [17]

    Automatic control of cardiac ablation catheter with deep reinforcement learning method,

    H. You, E. Bae, Y . Moon, J. Kweon, and J. Choi, “Automatic control of cardiac ablation catheter with deep reinforcement learning method,” Journal of Mechanical Science and Technology , vol. 33, no. 11, pp. 5415–5423, Nov 2019. [Online]. Available: https: //doi.org/10.1007/s12206-019-1036-0

    work page doi:10.1007/s12206-019-1036-0 2019

  18. [18]

    Deep reinforcement learning in medical imaging: A literature review,

    S. K. Zhou, H. N. Le, K. Luu, H. V Nguyen, and N. Ayache, “Deep reinforcement learning in medical imaging: A literature review,” Medical Image Analysis , vol. 73, p. 102193, 2021. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1361841521002395

    work page 2021

  19. [19]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P.Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention (MICCAI) , ser. LNCS, vol. 9351. Springer, 2015, pp. 234–241, (available on arXiv:1505.04597 [cs.CV]). [Online]. Available: http://lmb.informatik.uni-freiburg.de/Publications/ 2015/RFB15a

    Pith/arXiv arXiv 2015

  20. [20]

    Optimizing ensemble u-net architectures for robust coronary vessel segmentation in angiographic images,

    S.-S. Chang, C.-T. Lin, W.-C. Wang, K.-C. Hsu, Y .-L. Wu, C.-H. Liu, and Y . C. Fann, “Optimizing ensemble u-net architectures for robust coronary vessel segmentation in angiographic images,” Scientific Reports, vol. 14, no. 1, p. 6640, Mar 2024. [Online]. Available: https://doi.org/10.1038/s41598-024-57198-5

    doi:10.1038/s41598-024-57198-5 2024

  21. [21]

    Rapid multi- catheter segmentation for magnetic resonance image-guided catheter- based interventions,

    A. M. Aleong, A. Berlin, J. Borg, J. Helou, A. Beiki-Ardakani, A. Rink, S. Raman, P. Chung, and R. A. Weersink, “Rapid multi- catheter segmentation for magnetic resonance image-guided catheter- based interventions,” Medical Physics , 2024. [Online]. Available: https://aapm.onlinelibrary.wiley.com/doi/abs/10.1002/mp.17117

    work page doi:10.1002/mp.17117 2024

  22. [22]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” CoRR, vol. abs/1411.4038, 2014. [Online]. Available: http://arxiv.org/abs/1411.4038

    Pith/arXiv arXiv 2014

  23. [23]

    Efficient catheter segmentation in 3d cardiac ultrasound using slice-based fcn with deep supervision and f-score loss,

    H. Yang, C. Shan, A. F. Kolen, and P. H. N. de With, “Efficient catheter segmentation in 3d cardiac ultrasound using slice-based fcn with deep supervision and f-score loss,” in 2019 IEEE International Conference on Image Processing (ICIP) , 2019, pp. 260–264

    work page 2019

  24. [24]

    Deep high- resolution representation learning for visual recognition,

    J. Wang, K. Sun, T. Cheng, B. Jiang, C. Deng, Y . Zhao, D. Liu, Y . Mu, M. Tan, X. Wang, W. Liu, and B. Xiao, “Deep high- resolution representation learning for visual recognition,” CoRR, vol. abs/1908.07919, 2019. [Online]. Available: http://arxiv.org/abs/1908. 07919

    Pith/arXiv arXiv 1908

  25. [25]

    High-resolution feature based central venous catheter tip detection network in x-ray images,

    Y . Wang, H. K. Lam, Z.-G. Hou, R.-Q. Li, X.-L. Xie, and S.-Q. Liu, “High-resolution feature based central venous catheter tip detection network in x-ray images,” Med Image Anal , vol. 88, p. 102876, Jun. 2023

    work page 2023

  26. [26]

    Rethinking atrous convolution for semantic image segmentation,

    L. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethinking atrous convolution for semantic image segmentation,” CoRR, vol. abs/1706.05587, 2017. [Online]. Available: http://arxiv.org/abs/1706. 05587

    Pith/arXiv arXiv 2017

  27. [27]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” CoRR, vol. abs/1706.03762, 2017. [Online]. Available: http://arxiv.org/abs/1706.03762

    Pith/arXiv arXiv 2017

  28. [28]

    Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers,

    S. Zheng, J. Lu, H. Zhao, X. Zhu, Z. Luo, Y . Wang, Y . Fu, J. Feng, T. Xiang, P. H. Torr, and L. Zhang, “Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 6877–6886

    work page 2021

  29. [29]

    Segmenter: Trans- former for semantic segmentation,

    R. Strudel, R. Garcia, I. Laptev, and C. Schmid, “Segmenter: Trans- former for semantic segmentation,” arXiv preprint arXiv:2105.05633 , 2021

    Pith/arXiv arXiv 2021

  30. [30]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll´ar, and R. Girshick, “Segment anything,” arXiv:2304.02643, 2023

    Pith/arXiv arXiv 2023

  31. [31]

    Per-pixel classification is not all you need for semantic segmentation,

    B. Cheng, A. G. Schwing, and A. Kirillov, “Per-pixel classification is not all you need for semantic segmentation,” arXiv, 2021

    work page 2021

  32. [32]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. ´Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,” CoRR, vol. abs/2105.15203, 2021. [Online]. Available: https://arxiv.org/abs/2105.15203

    Pith/arXiv arXiv 2021

  33. [33]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” CoRR, vol. abs/2103.14030, 2021. [Online]. Available: https://arxiv.org/abs/2103.14030

    Pith/arXiv arXiv 2021

  34. [34]

    Fs-net: An encoder-decoder architecture for catheter segmentation and contact forces estimation in intracardiac catheters,

    P. Fekri, M. Zadeh, and J. Dargahi, “Fs-net: An encoder-decoder architecture for catheter segmentation and contact forces estimation in intracardiac catheters,” in 2025 IEEE World Haptics Conference (WHC), 2025, pp. 123–132

    work page 2025

  35. [35]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” CoRR, vol. abs/2005.12872, 2020. [Online]. Available: https://arxiv.org/abs/ 2005.12872

    Pith/arXiv arXiv 2005

  36. [36]

    Simple open-vocabulary object detection with vision transformers,

    M. Minderer, A. Gritsenko, A. Stone, M. Neumann, D. Weissenborn, A. Dosovitskiy, A. Mahendran, A. Arnab, M. Dehghani, Z. Shen, X. Wang, X. Zhai, T. Kipf, and N. Houlsby, “Simple open-vocabulary object detection with vision transformers,” 2022. [Online]. Available: https://arxiv.org/abs/2205.06230

    Pith/arXiv arXiv 2022

  37. [37]

    Tfs-vit: Token-level feature stylization for domain generalization,

    M. Noori, M. Cheraghalikhani, A. Bahri, G. A. Vargas Hakim, D. Osowiechi, I. B. Ayed, and C. Desrosiers, “Tfs-vit: Token-level feature stylization for domain generalization,” Pattern Recognition, vol. 149, p. 110213, 2024. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S003132032300910X

    work page 2024

  38. [38]

    Rec- ttt: Contrastive feature reconstruction for test-time training,

    M. Colussi, S. Mascetti, J. Dolz, and C. Desrosiers, “Rec- ttt: Contrastive feature reconstruction for test-time training,” in 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , 2025, pp. 6699–6708

    work page 2025

  39. [39]

    Watt: Weight average test-time adaptation of clip,

    D. Osowiechi, M. Noori, G. A. V . Hakim, M. Yazdanpanah, A. Bahri, M. Cheraghalikhani, S. Dastani, F. Beizaee, I. B. Ayed, and C. Desrosiers, “Watt: Weight average test-time adaptation of clip,”

    work page

  40. [2024]

    [Online]. Available: https://arxiv.org/abs/2406.13875 FEKRI et al.: TRANSFORCEG: A MULTI-MODAL ENCODER-DECODER ARCHITECTURE FOR STEREO SEGMENTATION AND 3D FORCE ESTIMATION 11 RGB-Side RGB-Top XRay1-Side XRay1-Top XRay2-Side XRay2-Top Normal segmentation Defocus segmentation Gaussian segmentation Stripe segmentation Fig. 4: Qualitative results across 3 out...

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.