Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Dynamic Robot-Assisted Surgery with Hierarchical Class-Incremental Semantic Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Hierarchical hyperbolic class encoding makes replay-free continual surgical segmentation work: TOPICS+ beats prior methods on six settings.

desk verdict A practical, well-benchmarked incremental extension of TOPICS for surgical CISS, with one unvalidated taxonomy that should be checked before trusting the gain magnitude. read the letter →

arxiv 2508.01713 v2 pith:WJ5Y6TQJ submitted 2025-08-03 cs.CV cs.AIcs.RO

Julia Hindel , Ema Mekic , Enamundram Naga Karthik , Rohit Mohan , Daniele Cattaneo , Maria Kalweit , Abhinav Valada This is my paper
classification cs.CVcs.AIcs.RO
keywords continuallearningclass-incrementalsemanticsegmentationhyperbolicspacePoincaréballhierarchicalpseudo-labelingDicelossroboticsurgerysyntheticdataset
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 asks whether a replay-free class-incremental segmentation model can keep adapting a surgical scene-understanding system to new instruments, tissues, and room equipment without storing or rehearsing old images. It argues yes, provided the classes are organized into a semantic tree and encoded in hyperbolic space. The proposed method, TOPICS+, adds a hierarchical Dice loss and hierarchical pseudo-labeling to the existing TOPICS approach, and it reports higher mean IoU than TOPICS and non-hierarchical baselines on six continual settings across Endovis18, MM-OR, and Syn-Mediverse. The clearest margin is a 4.4 percentage point gain over the best non-hierarchical methods on the refinement setting of a newly released 144-class refinement of Syn-Mediverse. If the claim holds, surgical robots could acquire new visual categories from new labels alone, which matters in clinical settings where storing old patient data is restricted.

What carries the argument

The load-bearing object is the Poincaré ball model of hyperbolic space with constant negative curvature, in which each semantic class is a hyperplane and class likelihood comes from signed distance to that hyperplane. A class taxonomy is encoded by treating leaf nodes and all their ancestors as separate output classes, so the loss can combine ancestor and descendant logits. TOPICS+ adds two components to this machinery: a hierarchical Dice loss that integrates the dense Dice loss at multiple levels of the tree to counter strong class imbalance, and hierarchical pseudo-labeling with per-level confidence thresholds that assigns old-class pixels to the most specific confident descendant rather than accepting any confident leaf prediction.

What would settle it

Retrain TOPICS+ on the refinement Syn-Mediverse setting with the class hierarchy randomly permuted but kept at the same tree shape; if mean IoU stays near the flat-baseline results, the semantic hierarchy is not load-bearing.

Watch

Extended reading notes

Core claim

In the paper's own terms, the central discovery is that hierarchical, hyperbolic class encoding is a viable replay-free route for continual semantic segmentation in robot-assisted surgery. TOPICS+ represents each class as a hyperplane in a Poincaré ball, enforces the semantic taxonomy through ancestor and descendant logits, and trains with a hierarchical Dice loss that handles class imbalance and penalizes high scores for sibling classes. Hierarchical pseudo-labeling assigns old-class pixels to the most specific descendant whose level-specific confidence threshold is exceeded, which suppresses the false-negative background predictions that hurt TOPICS in surgery scenes. Across six settings covering both disjoint class additions and refinements on Endovis18, MM-OR, and Syn-Mediverse, TOPICS+ outperforms TOPICS in every case and beats non-hierarchical baselines by up to 4.4 percentage points on the refinement Syn-Mediverse setting while retaining base-class knowledge.

Load-bearing premise

The method assumes that surgical scene classes form a meaningful tree-like hierarchy, that this hierarchy can be generated automatically by a single prompt to GPT-4o, and that encoding this particular tree in hyperbolic space is what drives the gains.

Editorial extensions

If this is right

  • A surgical robot could be updated with new instrument and anatomy classes from new labeled images alone, since TOPICS+ demonstrates replay-free class-incremental segmentation across six multi-task settings.
  • The benefit of hierarchical encoding grows in class-rich refinement settings: on the 144-class Syn-Mediverse refinement, TOPICS+ beats non-hierarchical methods by 4.4 percentage points while holding base-class performance.
  • False-negative background predictions, a known failure when instruments sit close to background tissue, are reduced by hierarchical pseudo-labeling with level-specific confidence thresholds.
  • The released 144-class Syn-Mediverse refinement provides a public benchmark on which future replay-free surgical class-incremental segmentation methods can be compared.

Reading between the lines

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

  • Editorial inference: the same hierarchical Dice loss and pseudo-labeling recipe should transfer to non-surgical continual segmentation wherever a clean class taxonomy exists, since neither component is specific to surgery.
  • Editorial inference: because the location-based hierarchy (TOPICS+ local) underperforms, the taxonomy's semantics matter more than its tree shape; testing automatically generated hierarchies against expert-curated ones would separate taxonomy quality from hyperbolic encoding.
  • Editorial inference: the per-level pseudo-labeling thresholds are set empirically, so an adaptive threshold schedule per task or per class frequency is a natural extension that could stabilize performance across unseen operating-room layouts.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes TOPICS+, a replay-free class-incremental semantic segmentation (CISS) method for robot-assisted surgery, extending the TOPICS framework. The method adds a hierarchical Dice loss, hierarchical pseudo-labeling, and GPT-4o-generated class taxonomies, and introduces six CISS settings across Endovis18, MM-OR, and Syn-Mediverse, plus a refined Syn-Mediverse label set with more than 144 classes. Experiments report that TOPICS+ outperforms TOPICS and several non-hierarchical baselines (PLOP, MiB, MiB+AWT, DKD) on both disjoint and refinement settings in terms of base and novel class mIoU.

Significance. If the claims are robust, the paper makes a useful contribution to continual learning in surgical scene understanding by combining hierarchical hyperbolic embeddings with domain-specific losses and releasing benchmarks and code. The strengths include comparison with published baseline implementations, ablations of the main components (curvature, hierarchical Dice loss, hierarchical pseudo-labeling), and a public online benchmark for Syn-Mediverse. However, the significance is tempered by the absence of multi-seed or significance-test evidence, the reliance on a single automatically generated taxonomy without validation, and incomplete details for reproducing the benchmark construction and the location-based hierarchy ablation.

major comments (4)
  1. [Section 3, Tables 1 and 2] All mIoU values in Tables 1 and 2 appear to come from a single training run, with no standard deviation, number of seeds, or significance tests. Since the central claim is that TOPICS+ outperforms all baselines, and several margins (e.g., versus TOPICS on the Endovis18 settings) are small, the reader cannot assess whether the observed differences are stable. Please report results over at least three seeds as mean +/- std, and perform a significance test (e.g., paired bootstrap or Wilcoxon) for the main comparisons, or justify why single-run comparisons are sufficient for these benchmark settings.
  2. [Section 2.3, Semantic Taxonomy Generation] The entire hierarchical loss and hierarchical pseudo-labeling scheme depend on the specific GPT-4o-generated class tree, but the paper does not validate that tree against a clinician-defined ontology, does not test alternative prompts, and does not perturb the tree structure. The location-based hierarchy comparison in Fig. 4 shows that the tree matters for performance, but it does not establish that the chosen GPT-4o tree is appropriate or that the claimed gains are not artifacts of one favorable taxonomy. Please add sensitivity experiments with alternative prompts, a manually constructed expert tree, and random tree perturbations, and report the effect on the main results in Tables 1 and 2.
  3. [Section 3.1, Implementation Details] The hierarchical pseudo-label thresholds (c0, c1, c2) and the hierarchical Dice loss coefficients are fixed to empirically chosen values, but no sensitivity analysis is provided. Because these thresholds determine when pixels are assigned to leaf versus ancestor classes, and the loss coefficients weigh the hierarchical terms, the reported improvements could be sensitive to these choices. Please include a sweep over thresholds and loss coefficients on at least one dataset, or provide a justification for the selected values.
  4. [Section 2.3 and Fig. 4, location-based hierarchy] The text states that graphs linking objects that frequently appear in close proximity are investigated in Sec. 3.2, but Section 3.2 is the qualitative results section and contains no description of how the location-based hierarchy (TOPICS+ local) was constructed. This makes the location-based ablation in Fig. 4 non-reproducible and its interpretation unclear. Please move the construction details for the location-based hierarchy to Section 2.3 or add a dedicated ablation subsection describing the graph construction and the resulting tree.
minor comments (5)
  1. [Section 2.1, Syn-Mediverse label refinement] The paper announces 'more than 144 classes' and 'over 144 fine categories' but does not state the exact number of final classes, the precise merge and filter rules, or the class distribution in the released labels. Since the refined label set is a core contribution and is hosted online, these statistics are necessary for users and for interpreting the difficulty of the incremental setting.
  2. [Section 3.1, Implementation Details] Several hyperparameter values in the implementation details are not legible in the submitted manuscript (e.g., the hierarchical Dice loss coefficients and the pseudo-label thresholds appear as placeholder symbols). Please ensure all equations and numerical values are typeset correctly so that the experimental setup is reproducible.
  3. [References] Some references are incomplete: reference [14] lacks a publication year and reference [27] lacks a year and venue. Please update the bibliography to include all required metadata.
  4. [Conclusion and Discussion] The paper has no explicit limitations section. Given the reliance on a single GPT-4o-generated taxonomy and the lack of multi-seed evaluations, a short limitations discussion would help readers calibrate the scope of the claims.
  5. [Section 1, Introduction] The sentence 'we emphasize that incremental learning scenarios should not be limited to specific categories' is grammatically awkward; consider rephrasing for clarity.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the central claim rests on external baselines and public datasets, with only minor non-load-bearing self-citations.

full rationale

The paper's central claim is that TOPICS+ outperforms state-of-the-art CISS methods on six continual surgical-segmentation settings. This is an empirical claim supported by experiments against PLOP, MiB, MiB+AWT, and DKD using their published code, and by evaluation on public Endovis18 and MM-OR datasets. The main self-citations are to TOPICS [10], the authors' own predecessor, and to Syn-Mediverse [18], also from the same group. However, TOPICS is not invoked as an unverified uniqueness theorem or as the sole justification for the hierarchy; it is re-run as a baseline, and the benefit of hierarchical encoding is tested against non-hierarchical methods and against a location-based hierarchy in Fig. 4. The Syn-Mediverse refined labels and online benchmark are new data contributions rather than components of the loss or the evaluation protocol that reduce to the method's own outputs. The GPT-4o-generated taxonomy is a load-bearing input, and its sensitivity to prompt and tree choice is not tested, but that is an external-validity and correctness concern, not a circularity: the taxonomy is not defined in terms of the method's predictions, and the reported gains are not forced by construction. The hierarchical pseudo-labeling thresholds are tuned on validation data, which is standard model selection rather than a fitted parameter being renamed as a prediction. Overall, no step in the derivation chain equates a claimed prediction to its own input, so the paper is not significantly circular.

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

The central claim depends on four groups of free choices: pseudo-label thresholds, loss weights, curvature, and the GPT-generated taxonomy. These are not derived from data or theory, and several are explicitly tuned on validation performance. The axioms are standard for CISS evaluation but are assumed rather than demonstrated. No new physical, geometric, or algorithmic entities are invented beyond the method itself.

free parameters (4)
  • Hierarchical pseudo-label thresholds c0, c1, c2 = 0.9, 0.8, 0.7
    Chosen empirically based on validation performance (Section 3.1); not derived from theory and could change results across datasets.
  • Loss weight coefficients in hierarchical dice loss = Not fully recoverable in extracted text
    The hierarchical loss introduces weighting terms for Dice, relation, and CE terms (Section 2.2) that are set to fixed values in experiments; these weights directly affect the reported mIoU.
  • Poincaré ball curvature c = -1
    Fixed at -1 for main experiments and ablated in Table 3; the curvature controls hyperbolic distances and is selected without a principled rule.
  • GPT-4o taxonomy generation prompt = Prompt quoted in Section 2.3, no temperature or decoding details
    The semantic hierarchy is generated by an LLM with a single prompt; the resulting tree is a free design choice that affects all hierarchical losses and pseudo-labeling.
assumptions (4)
  • domain assumption Surgical scene classes admit a meaningful tree-like semantic hierarchy that can be generated automatically and encoded in hyperbolic space.
    Section 2.3 constructs GPT-4o-generated taxonomies and Section 2.2 assumes hyperplanes in the Poincaré ball can reflect this hierarchy, but the hierarchy is not validated against expert ontologies or tested for prompt sensitivity.
  • domain assumption The CISS protocol with overlapping images and background shift is an appropriate surrogate for real continual deployment in surgery.
    Section 2.1 defines tasks with overlapping images where only current-class pixels are labeled; this standard setup assumes that seeing future-class pixels as background does not undermine the evaluation.
  • domain assumption DeepLabV3 with ResNet-101 provides a suitable pretrained base for incremental segmentation.
    Section 2.2 adopts the architecture without comparing across backbones or discussing how pretrained features interact with hyperbolic projection.
  • domain assumption mIoU on the proposed test splits is a sufficient measure of continual learning performance.
    The paper reports mean IoU only, without variance, forgetting metrics, or task-boundary sensitivity analyses, yet bases its main conclusions on mIoU rankings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic Robot-Assisted Surgery with Hierarchical Class-Incremental Semantic Segmentation." pith.science (2026). https://pith.science/paper/WJ5Y6TQJ

@misc{pith2026250801713,
  author       = {Pith},
  title        = {Pith review of: Dynamic Robot-Assisted Surgery with Hierarchical Class-Incremental Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJ5Y6TQJ}},
  note         = {Machine review of arXiv:2508.01713}
}
read the original abstract

Robot-assisted surgeries rely on accurate and real-time scene understanding to safely guide surgical instruments. However, segmentation models trained on static datasets face key limitations when deployed in these dynamic and evolving surgical environments. Class-incremental semantic segmentation (CISS) allows models to continually adapt to new classes while avoiding catastrophic forgetting of prior knowledge, without training on previous data. In this work, we build upon the recently introduced Taxonomy-Oriented Poincar\'e-regularized Incremental Class Segmentation (TOPICS) approach and propose an enhanced variant, termed TOPICS+, specifically tailored for robust segmentation of surgical scenes. Concretely, we incorporate the Dice loss into the hierarchical loss formulation to handle strong class imbalances, introduce hierarchical pseudo-labeling, and design tailored label taxonomies for robotic surgery environments. We also propose six novel CISS benchmarks designed for robotic surgery environments including multiple incremental steps and several semantic categories to emulate realistic class-incremental settings in surgical environments. In addition, we introduce a refined set of labels with more than 144 classes on the Syn-Mediverse synthetic dataset, hosted online as an evaluation benchmark. We make the code and trained models publicly available at http://topics.cs.uni-freiburg.de.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 27 canonical work pages

  1. [1]

    arXiv preprint arXiv:1902.06426 (2019)

    Allan, M., et al.: 2017 robotic instrument segmentation challenge. arXiv preprint arXiv:1902.06426 (2019)

    arXiv 2019

  2. [2]

    arXiv preprint arXiv:2001.11190 (2020)

    Allan, M., et al.: 2018 robotic scene segmentation challenge. arXiv preprint arXiv:2001.11190 (2020)

    arXiv 2020

  3. [3]

    Baek, D., Oh, Y., Lee, S., Lee, J., Ham, B.: Decomposed knowledge distillation for class-incremental semantic segmentation. In: Adv. Neural Inform. Process. Syst. pp. 10380–10392 (2022)

    work page 2022

  4. [4]

    In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2023

    Bai, L., Islam, M., Ren, H.: Revisiting distillation for continual learning on visual question localized-answering in robotic surgery. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2023. pp. 68–78 (2023)

    work page 2023

  5. [5]

    Biomedical Engineering Advances6, 100109 (2023)

    Biswas, P., Sikander, S., Kulkarni, P.: Recent advances in robot-assisted surgical systems. Biomedical Engineering Advances6, 100109 (2023)

    work page 2023

  6. [6]

    In: Proc

    Cermelli, F., Mancini, M., Rota Bulò, S., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: Proc. IEEE Conf. Comput. Vis. Pattern Recog. (2020)

    work page 2020

  7. [7]

    Douillard, A., Chen, Y., Dapogny, A., Cord, M.: Plop: Learning without forgetting for continual semantic segmentation. Proc. IEEE Conf. Comput. Vis. Pattern Recog. pp. 4039–4049 (2020)

    work page 2020

  8. [8]

    Ganea, O., Bécigneul, G., Hofmann, T.: Hyperbolic neural networks. Adv. Neural Inform. Process. Syst.31 (2018)

    work page 2018

Show all 31 references
  1. [9]

    In: IEEE/CVF Winter Conf

    Goswami, D., Schuster, R., van de Weijer, J., Stricker, D.: Attribution-aware weight transfer: A warm-start initialization for class-incremental semantic segmentation. In: IEEE/CVF Winter Conf. on Applications of Comput. Vis. pp. 3195–3204 (2023) 10 Hindel et al

    2023

  2. [10]

    IEEE Robotics and Automation Letters 10(2), 1904–1911 (2025)

    Hindel, J., Cattaneo, D., Valada, A.: Taxonomy-aware continual semantic segmenta- tion in hyperbolic spaces for open-world perception. IEEE Robotics and Automation Letters 10(2), 1904–1911 (2025)

    2025

  3. [11]

    In: Proc

    Ji, Z., Guo, D., Wang, P., Yan, K., Lu, L., Xu, M., Wang, Q., Ge, J., Gao, M., Ye, X., Jin, D.: Continual segment: Towards a single, unified and non-forgetting continual segmentation model of 143 whole-body organs in ct scans. In: Proc. Int. Conf. Comput. Vis. pp. 21083–21094 (2023)

    2023

  4. [12]

    Karthik, E.N., Kerbrat, A., Labauge, P., Granberg, T., Talbott, J., Reich, D.S., Filippi, M., Bakshi, R., Callot, V., Chandar, S., et al.: Segmentation of multiple sclerosis lesions across hospitals: Learn continually or train from scratch? arXiv preprint arXiv:2210.15091 (2022)

    2022 arXiv

  5. [13]

    In: Leibe, B., Matas, J., Sebe, N., Welling, M

    Li, Z., Hoiem, D.: Learning without forgetting. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Proc. Springer Eur. Conf. Comput. Vis. pp. 614–629 (2016)

    2016

  6. [14]

    arXiv preprint arXiv:2203.02100

    Liu, P., Wang, X., Fan, M., Pan, H., Yin, M., Zhu, X., Du, D., Zhao, X., Xiao, L., Ding, L., Wu, X., Zhou, S.K.: Learning incrementally to segment multiple organs in a ct image. arXiv preprint arXiv:2203.02100

    arXiv

  7. [15]

    In: Medical Image Computing and Computer Assisted Intervention

    Liu, X., Shih, H.A., Xing, F., Santarnecchi, E., El Fakhri, G., Woo, J.: Incremental learning for heterogeneous structure segmentation in brain tumor mri. In: Medical Image Computing and Computer Assisted Intervention. pp. 46–56 (2023)

    2023

  8. [16]

    In: Psychology of learning and motivation, vol

    McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. In: Psychology of learning and motivation, vol. 24, pp. 109–165. Elsevier (1989)

    1989

  9. [17]

    In: International conference on 3D vision

    Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: International conference on 3D vision. pp. 565–571 (2016)

    2016

  10. [18]

    IEEE Robotics and Automation Letters9(8), 7094–7101 (2024)

    Mohan, R., Arce, J., Mokhtar, S., Cattaneo, D., Valada, A.: Syn-mediverse: A mul- timodal synthetic dataset for intelligent scene understanding of healthcare facilities. IEEE Robotics and Automation Letters9(8), 7094–7101 (2024)

    2024

  11. [19]

    arXiv preprint arXiv:2303.08774 (2024)

    OpenAI, et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2024)

    2024 arXiv

  12. [20]

    In: Medical Image Computing and Computer Assisted Intervention

    Ozdemir, F., Fuernstahl, P., Goksel, O.: Learn the new, keep the old: Extending pretrained models with new anatomy and images. In: Medical Image Computing and Computer Assisted Intervention. pp. 361–369 (2018)

    2018

  13. [21]

    International Journal of Computer Assisted Radiology and Surgery 14(7), 1187–1195 (May 2019)

    Ozdemir, F., Goksel, O.: Extending pretrained segmentation networks with addi- tional anatomical structures. International Journal of Computer Assisted Radiology and Surgery 14(7), 1187–1195 (May 2019)

    2019

  14. [22]

    Electronics 13(23) (2024)

    Picozzi, P., Nocco, U., Labate, C., Gambini, I., Puleo, G., Silvi, F., Pezzillo, A., Mantione, R., Cimolin, V.: Advances in robotic surgery: A review of new surgical platforms. Electronics 13(23) (2024)

    2024

  15. [23]

    arXiv preprint arXiv:2405.13482 (2024)

    Qazi, M.A., Hashmi, A.U.R., Sanjeev, S., Almakky, I., Saeed, N., Gonzalez, C., Yaqub, M.: Continual learning in medical imaging: A survey and practical analysis. arXiv preprint arXiv:2405.13482 (2024)

    2024 arXiv

  16. [24]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Roy, K., Moghadam, P., Harandi, M.: L3dmc: Lifelong learning using distillation via mixed-curvature space. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 123–133 (2023)

    2023

  17. [25]

    In: IEEE/RSJ International conference on intelligent robots and systems (IROS) workshop, state estimation and terrain perception for all terrain mobile robots

    Valada, A., Dhall, A., Burgard, W.: Convoluted mixture of deep experts for robust semantic segmentation. In: IEEE/RSJ International conference on intelligent robots and systems (IROS) workshop, state estimation and terrain perception for all terrain mobile robots. vol. 2, p. 1 (2016)

    2016

  18. [26]

    In: Proc

    Vödisch, N., Cattaneo, D., Burgard, W., Valada, A.: Covio: Online continual learning for visual-inertial odometry. In: Proc. IEEE Conf. Comput. Vis. Pattern Recog. pp. 2464–2473 (2023) Hierarchical Class-Incremental Semantic Segmentation 11

    2023

  19. [27]

    arXiv preprint arXiv:2503.06598

    Xu, H., Xue, T., Liu, D., Chen, Y., Zhang, F., Westin, C., Kikinis, R., O’Donnell, L., Cai, W.: Multico3d: Multi-label voxel contrast for one-shot incremental segmentation of 3d neuroimages. arXiv preprint arXiv:2503.06598

    arXiv

  20. [28]

    IEEE Transactions on Medical Imaging43, 2291– 2302 (2024)

    Xu, M., Islam, M., Bai, L., Ren, H.: Privacy-preserving synthetic continual semantic segmentation for robotic surgery. IEEE Transactions on Medical Imaging43, 2291– 2302 (2024)

    2024

  21. [29]

    In: Proc

    Zhang, C.B., Xiao, J.W., Liu, X., Chen, Y.C., Cheng, M.M.: Representation compensation networks for continual semantic segmentation. In: Proc. IEEE Conf. Comput. Vis. Pattern Recog. pp. 7053–7064 (2022)

    2022

  22. [30]

    In: Medical Image Computing and Computer Assisted Intervention

    Zhang, Y., Li, X., Chen, H., Yuille, A.L., Liu, Y., Zhou, Z.: Continual learning for abdominal multi-organ and tumor segmentation. In: Medical Image Computing and Computer Assisted Intervention. pp. 35–45 (2023)

    2023

  23. [31]

    In: Proc

    Özsoy, E., Pellegrini, C., Czempiel, T., Tristram, F., Yuan, K., Bani-Harouni, D., Eck, U., Busam, B., Keicher, M., Navab, N.: Mm-or: A large multimodal operating room dataset for semantic understanding of high intensity surgical environments. In: Proc. IEEE Conf. Comput. Vis....

    2025

Pith tools

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