Pith. sign in

REVIEW 3 major objections 6 minor 43 references

CDPDNet: Integrating Text Guidance with Hybrid Vision Encoders for Medical Image Segmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single text-guided network can segment 25 organs, 6 tumors, and 1 cyst across 11 partially labeled CT datasets, reporting the highest average Dice (85.08%) across all compared methods.

desk verdict Solid multi-dataset segmentation architecture study; the unseen-dataset generalization numbers rest on manually chosen task IDs, so take those claims with salt. read the letter →

arxiv 2505.18958 v2 pith:TUJ3ZMJK submitted 2025-05-25 cs.CV

Jiong Wu , Yang Xing , Boxiao Yu , Wei Shao , Kuang Gong This is my paper · ORCID
classification cs.CV
keywords medicalimagesegmentationpartiallylabeleddatasetsCLIPtextembeddingsDINOv2text-basedtaskpromptmulti-organtumorCTimaging
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

CDPDNet is a dynamic-head segmentation model that combines a frozen DINOv2 vision transformer, a CNN encoder, and CLIP text embeddings to learn from 11 partially labeled CT datasets covering 25 organs, 6 tumors, and 1 cyst. The paper's central claim is that injecting text embeddings at multiple feature scales and generating a task-specific prompt from text lets one network separate tasks that share few labels and generalize to datasets it never trained on. Across the 11 training tasks it reports the highest average Dice similarity coefficient (85.08%) and the lowest average Hausdorff distance (10.62), a boundary-error metric, among six published baselines plus a CLIP-DoDNet variant. On two unseen datasets, AbdomenCTCT and 3D-IRCADb, it reports 77.40% average Dice versus 66.00% for the best baseline, without fine-tuning. A sympathetic reader would care because the partial-labeling problem is what forces most clinics to train separate models for each organ or tumor, and the paper argues a single text-prompted model can replace them.

What carries the argument

The load-bearing object is the Text-based Task Prompt Generation (TTPG) module, which builds a task-specific spatial map of shape $B \times N \times h \times w \times d$ and selects the $i$-th channel according to the task ID for injection into the mask decoder. Around that object sit three supporting mechanisms: the alignment function $\psi(v,t)$ that projects CLIP text embeddings into visual space and applies them as a Hadamard product with multi-scale visual features; four cross-attention modules that treat DINOv2 dense features as queries and CNN features as keys and values, fusing global context with local detail; and a 3D adaptor that stacks axial information from DINOv2's 2D slice features so the frozen vision transformer can serve 3D volumes. Together these mechanisms carry the claim that text guidance at multiple scales, not just the bottleneck, explains the model's separation of partially labeled tasks and its transfer to unseen datasets.

What would settle it

Run CDPDNet on an unseen CT dataset while selecting task IDs without using knowledge of the target structures (for example, random IDs or IDs chosen by an automatic text-matching rule), then compare average Dice with the 77.40% reported for manually selected IDs; if the average falls toward the 66.00% baseline, the generalization claim is an artifact of task-ID selection.

Watch

Extended reading notes

Core claim

The discovery the paper is trying to establish is that task-aware text guidance, applied at multiple resolutions rather than only at the bottleneck, is what makes a single dynamic segmentation network succeed on partially labeled data. Concretely, CLIP text embeddings for organ and tumor names are projected into the visual space through the alignment function $\psi(v,t) = (W_a t + b_a) \odot v + W_b t + b_b$ at four feature levels, and a Text-based Task Prompt Generation module fuses the general text embedding, the task embedding, and bottleneck features into a task-specific spatial map selected by the task ID. The paper reports this configuration achieves 85.08% average Dice across 11 tasks and 77.40% on unseen scans, with its largest margins on small and pathological structures such as kidney tumor, lung tumor, colon tumor, gallbladder, and left adrenal gland. Stated on the paper's own terms, the text prompt does not just name the target; it steers the decoder's dynamic parameters so that one model can serve all tasks.

Load-bearing premise

The reported gains on unseen datasets depend on a human choosing the right task ID for each new dataset, and the paper's own ablation shows that choice moves Dice from about 60% to 90%, so the 77.40% generalization number is only as strong as that manual selection step.

Editorial extensions

If this is right

  • A single trained CDPDNet can output segmentations for 25 organs, 6 tumors, and 1 cyst across 11 datasets, so clinics would no longer need a separate model per organ or tumor.
  • Changing the task ID changes the generated text prompt and therefore the decoder's behavior, giving a concrete steering mechanism for user-interactive segmentation.
  • Because the DINOv2 and CLIP encoders stay frozen, scaling to more tasks adds trainable parameters only in the CNN branch, cross-attention modules, alignment function, and decoder.
  • On two unseen datasets the model improves average Dice by 11.4 points over the best baseline without fine-tuning, suggesting text-guided dynamic models can transfer across CT protocols better than vision-only dynamic models.

Reading between the lines

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

  • An automatic task-ID selector, built by matching CLIP text embeddings of the target structures to the training task whose labels overlap most, would remove the manual step the paper flags as its main limitation; the paper's own Fig. 6 shows Dice ranging from roughly 60% to 90% depending on which ID is chosen.
  • The multi-scale alignment design predicts a specific ablation pattern: removing the early-resolution alignments should hurt small-structure Dice more than removing the bottleneck alignment, because the paper's argument is that low-level features carry fine boundary information earlier models missed.
  • Because the prompt template names the imaging modality ('computerized tomography'), adapting the architecture to MRI or ultrasound may only require changing that template and re-embedding the structures; the paper does not test this.
  • For a fair comparison on unseen data, the full distribution of Dice across task IDs matters more than the single 77.40% average, and the paper reports that distribution only for AbdomenCTCT, so the generalization number is conditional on ID selection.
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

3 major / 6 minor

Summary. The manuscript proposes CDPDNet, a unified medical image segmentation model that combines a frozen DINOv2 vision transformer, a CNN encoder, CLIP text embeddings, and a Text-based Task Prompt Generation (TTPG) module to address partially labeled multi-organ and tumor segmentation. The model is evaluated on 11 CT datasets covering 32 ROIs and, for generalization, on two unseen datasets (AbdomenCTCT and 3D-IRCADb). The authors report the highest average DSC (85.08%) and lowest average HD (10.62) across the 11 tasks, as well as strong results on unseen datasets (77.40% average DSC versus 66.00% for the best baseline), and provide code and pretrained models.

Significance. If the claims are substantiated, CDPDNet would be a useful contribution to partially labeled medical segmentation: the hybrid DINOv2/CNN encoder with multi-scale CLIP text alignment is a sensible architecture, the benchmark is unusually broad (11 datasets, 32 ROIs, consistent training settings), and the ablations decompose the contributions of the main modules. The availability of code and pretrained models is a concrete strength. However, the headline generalization claim is weakened by the manual post hoc selection of task IDs on unseen datasets, which the paper itself acknowledges; as presented, that result is an upper bound rather than an out-of-the-box generalization evaluation.

major comments (3)
  1. [Sec. IV-C, Sec. IV-D2, Fig. 6, Sec. V] The central generalization claim in Table IV (77.40% average DSC versus 66.00% for the best baseline) is not an out-of-the-box evaluation, because the task ID for each unseen dataset is selected manually after seeing the target data. Section IV-D2 and Fig. 6 show that on AbdomenCTCT the reported DSC varies from roughly 60% to 90% depending on task ID, and Section V explicitly concedes that task ID selection is manual. The manuscript does not state which task IDs were used for Table IV or how a user would choose them in practice, so the comparison effectively involves test-time selection; the 11.4-point margin over the best baseline is therefore not a predictive generalization claim. Please provide an automatic task-ID selection mechanism, or report results with a fixed protocol (e.g., averaged over all task IDs or chosen by a principled rule), and restrict the abstract and conclusion claims accordingly.
  2. [Abstract, Sec. IV-B, Table II] The abstract's assertion that CDPDNet 'consistently outperformed existing state-of-the-art segmentation methods' is contradicted by the paper's own results: in Table II, on WORD, CDPDNet has DSC 89.06% versus PromptUniseg's 89.86%, and the text in Sec. IV-B acknowledges the exception. Please replace 'consistently outperformed' with a precise statement, such as 'achieved the best average DSC and HD across the 11 tasks, with task-by-task exceptions.'
  3. [Tables II-IV, Sec. IV-B] All DSC/HD comparisons are reported as single numbers per task or ROI without error bars, confidence intervals, or significance tests. Several margins are small (e.g., CHAOS 82.88 versus 82.69 and AMOS22 84.35 versus 84.34 in Table II; LKid 94.23 versus 94.54 in Table III), so it is unclear which of the claimed improvements are robust. Please report variance across folds, per-case standard deviations, or statistical tests for the main comparisons.
minor comments (6)
  1. [Sec. III-C4, Eq. (6)] The alignment function is introduced as ψ: R^Ωt → R^Ωv, but Eq. (6) defines ψ(v,t) with both visual and text inputs; the domain and the notation should be reconciled.
  2. [Table IV] The baseline name is spelled 'CLIP-DODNet' in Table IV but 'CLIP-DoDNet' elsewhere; please unify the spelling.
  3. [Sec. IV-C] The generalization evaluation on 3D-IRCADb covers only the liver tumor, which is a narrow test of cross-dataset generalization; this should be acknowledged when interpreting the aggregated results in Table IV.
  4. [Table III] Several cells are marked 'N/A' without a footnote explaining whether the metric was undefined, not computed, or not applicable for that method and ROI; please add a legend.
  5. [Fig. 3] The caption refers to panels '(a)' and '(b)', but the figure as presented appears to be a single panel; please align the caption with the actual layout.
  6. [Sec. IV-A2] The learning-rate schedule and warm-up are described, but the total number of training epochs is not stated; please add the epoch count.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-level circularity; the only flagged issue is post hoc task-ID selection on unseen datasets, which weakens the generalization claim but does not reduce the method to its inputs.

  1. other [Sec. IV-D2 and Sec. V]
    "Task ID is a critical hyperparameter in the proposed method, particularly when segmenting datasets different from those used during training. ... As shown in Fig. 6, segmentation accuracy varied significantly with task IDs. ... One limitation of this work is the need for manual selection of the task ID when applying the trained model to downstream segmentation tasks."

    The headline generalization result in Table IV ('77.40% versus 66.00%') is reported after choosing task IDs for the unseen datasets, and Fig. 6 shows that on AbdomenCTCT the same trained model ranges from roughly 60% to 90% DSC depending on the task ID. Since the paper admits the ID is manually selected, the unseen-dataset 'prediction' is conditional on a test-time choice rather than a fixed out-of-the-box rule. This is not a definitional equivalence, but it is the only load-bearing step that resembles fitting an input on the evaluated data; it does not affect Tables II/III or the ablations.

full rationale

The central contribution is an empirical architecture evaluation, not a derivation from first principles. CDPDNet combines frozen DINOv2 and CLIP encoders with trainable CNN, cross-attention, alignment, and TTPG modules under standard Dice and cross-entropy losses, and the 11-task results in Tables II/III are measured against six external baselines under identical settings. Ablations in Table V separate the contribution of each module, and the best configuration is not forced by any equation in the paper. There is no load-bearing self-citation chain: cited backbones and baselines come from external groups, and no uniqueness theorem or ansatz is imported from the authors' own prior work. The only flagged issue is the manual selection of task ID on the two unseen datasets, disclosed in Secs. IV-D2 and V; because this selection can swing per-organ DSC by tens of points, it weakens the generalization claim, but it is an evaluation-protocol limitation rather than a constructional circularity. The 11-task central comparison remains independent, so the overall circularity score is low.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities; the modules (3D adaptor, TTPG) are architectural components, not invented entities. The main free parameter is the manually chosen task ID used for unseen datasets.

free parameters (1)
  • Task ID for unseen datasets = 6, 8, 9, 10 (selected per dataset)
    The model requires a task ID at inference; for AbdomenCTCT and 3D-IRCADb the authors manually select IDs that yield high DSC. Fig. 6 shows DSC ranges from about 60% to 90% depending on ID, so the choice is a fitted hyperparameter affecting the reported 77.40% generalization average.
assumptions (3)
  • domain assumption DINOv2, pretrained on 2D natural images, provides features useful for 3D CT segmentation after slicing and a 3D adaptor.
    Used in Sec III-C1; the paper assumes the self-supervised features transfer across domains without an experiment isolating this transfer.
  • domain assumption CLIP text embeddings of organ/tumor names carry semantic relationships that align with CT visual features through a learned linear transform.
    Sec III-C4 relies on this; CLIP was trained on natural images and paired text, not medical images. The alignment is learned, but the assumption that the text space is informative for CT is untested independently.
  • domain assumption Combining datasets with masked backpropagation, where only annotated classes contribute to the loss, is sufficient to learn a unified segmentation model.
    Sec III-F; common practice from DoDNet and UniSeg, not re-derived here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CDPDNet: Integrating Text Guidance with Hybrid Vision Encoders for Medical Image Segmentation." pith.science (2026). https://pith.science/paper/TUJ3ZMJK

@misc{pith2026250518958,
  author       = {Pith},
  title        = {Pith review of: CDPDNet: Integrating Text Guidance with Hybrid Vision Encoders for Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TUJ3ZMJK}},
  note         = {Machine review of arXiv:2505.18958}
}
read the original abstract

Most publicly available medical segmentation datasets are only partially labeled, with annotations provided for a subset of anatomical structures. When multiple datasets are combined for training, this incomplete annotation poses challenges, as it limits the model's ability to learn shared anatomical representations among datasets. Furthermore, vision-only frameworks often fail to capture complex anatomical relationships and task-specific distinctions, leading to reduced segmentation accuracy and poor generalizability to unseen datasets. In this study, we proposed a novel CLIP-DINO Prompt-Driven Segmentation Network (CDPDNet), which combined a self-supervised vision transformer with CLIP-based text embedding and introduced task-specific text prompts to tackle these challenges. Specifically, the framework was constructed upon a convolutional neural network (CNN) and incorporated DINOv2 to extract both fine-grained and global visual features, which were then fused using a multi-head cross-attention module to overcome the limited long-range modeling capability of CNNs. In addition, CLIP-derived text embeddings were projected into the visual space to help model complex relationships among organs and tumors. To further address the partial label challenge and enhance inter-task discriminative capability, a Text-based Task Prompt Generation (TTPG) module that generated task-specific prompts was designed to guide the segmentation. Extensive experiments on multiple medical imaging datasets demonstrated that CDPDNet consistently outperformed existing state-of-the-art segmentation methods. Code and pretrained model are available at: https://github.com/wujiong-hub/CDPDNet.git.

Figures

Figures reproduced from arXiv: 2505.18958 by the authors.

Figure 1
Figure 1. Overview of the proposed CLIP-DINO-Prompt Driven segmentation network (CDPDNet). It comprised three main [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the proposed Text-based Task Prompt [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. (a) Training and testing image composition. (b) Annotated 25 organs, 6 tumors, and a kidney cyst for 11 different [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visual comparison of segmentation methods on 5 representative organ segmentation samples from the testing dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison of segmentation methods on 5 representative tumor segmentation samples from the testing dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Average DSC distribution for each of the 11 organs on [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 32 canonical work pages

  1. [1]

    Deep learning in medical imaging and radiation therapy,

    B. Sahiner, A. Pezeshk, L. M. Hadjiiski, X. Wang, K. Drukker, K. H. Cha, R. M. Summers, and M. L. Giger, “Deep learning in medical imaging and radiation therapy,”Medical physics, vol. 46, no. 1, pp. e1–e36, 2019

    work page 2019

  2. [2]

    Computer-aided detection of subsolid nodules at chest ct: improved performance with deep learning–based ct section thickness reduction,

    S. Park, S. M. Lee, W. Kim, H. Park, K.-H. Jung, K.-H. Do, and J. B. Seo, “Computer-aided detection of subsolid nodules at chest ct: improved performance with deep learning–based ct section thickness reduction,”Radiology, vol. 299, no. 1, pp. 211–219, 2021

    work page 2021

  3. [3]

    Automated deep learning-based detection and segmentation of lung tumors at ct,

    M. Kashyap, X. Wang, N. Panjwani, M. Hasan, Q. Zhang, C. Huang, K. Bush, A. Chin, L. K. Vitzthum, P. Donget al., “Automated deep learning-based detection and segmentation of lung tumors at ct,”Radi- ology, vol. 314, no. 1, p. e233029, 2025

    work page 2025

  4. [4]

    Crossbar-net: A novel convolutional neural network for kidney tumor segmentation in ct images,

    Q. Yu, Y . Shi, J. Sun, Y . Gao, J. Zhu, and Y . Dai, “Crossbar-net: A novel convolutional neural network for kidney tumor segmentation in ct images,”IEEE transactions on image processing, vol. 28, no. 8, pp. 4060–4074, 2019

    work page 2019

  5. [5]

    Light-weight hybrid convolutional network for liver tumor segmentation

    J. Zhang, Y . Xie, P. Zhang, H. Chen, Y . Xia, and C. Shen, “Light-weight hybrid convolutional network for liver tumor segmentation.” inIJCAI, vol. 19, 2019, pp. 4271–4277. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 10

    work page 2019

  6. [6]

    Multi-scale coarse-to-fine segmentation for screening pancreatic ductal adenocarci- noma,

    Z. Zhu, Y . Xia, L. Xie, E. K. Fishman, and A. L. Yuille, “Multi-scale coarse-to-fine segmentation for screening pancreatic ductal adenocarci- noma,” inMICCAI2019. Springer, 2019, pp. 3–12

    work page 2019

  7. [7]

    Learning from partially labeled data for multi-organ and tumor segmentation,

    Y . Xie, J. Zhang, Y . Xia, and C. Shen, “Learning from partially labeled data for multi-organ and tumor segmentation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 12, pp. 14 905– 14 919, 2023

    work page 2023

  8. [8]

    Med3d: Transfer learning for 3d medical image analysis,

    S. Chen, K. Ma, and Y . Zheng, “Med3d: Transfer learning for 3d medical image analysis,”arXiv preprint arXiv:1904.00625, 2019

    arXiv 1904

Show all 43 references
  1. [9]

    Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction,

    X. Fang and P. Yan, “Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction,”IEEE Transactions on Medical Imaging, vol. 39, no. 11, pp. 3619–3629, 2020

    2020

  2. [10]

    Marginal loss and exclusion loss for partially supervised multi-organ segmentation,

    G. Shi, L. Xiao, Y . Chen, and S. K. Zhou, “Marginal loss and exclusion loss for partially supervised multi-organ segmentation,”Medical Image Analysis, vol. 70, p. 101979, 2021

    2021

  3. [11]

    Dodnet: Learning to segment multi-organ and tumors from multiple partially labeled datasets,

    J. Zhang, Y . Xie, Y . Xia, and C. Shen, “Dodnet: Learning to segment multi-organ and tumors from multiple partially labeled datasets,” in CVPR, 2021, pp. 1195–1204

    2021

  4. [12]

    Uniseg: A prompt-driven universal segmentation model as well as a strong representation learner,

    Y . Ye, Y . Xie, J. Zhang, Z. Chen, and Y . Xia, “Uniseg: A prompt-driven universal segmentation model as well as a strong representation learner,” inMICCAI. Springer, 2023, pp. 508–518

    2023

  5. [13]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  6. [14]

    Clip-driven universal model for organ segmentation and tumor detection,

    J. Liu, Y . Zhang, J.-N. Chen, J. Xiao, Y . Lu, B. A Landman, Y . Yuan, A. Yuille, Y . Tang, and Z. Zhou, “Clip-driven universal model for organ segmentation and tumor detection,” inICCV, 2023, pp. 21 152–21 164

    2023

  7. [15]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”TMLR, 2023

    2023

  8. [16]

    Prior-aware neural network for partially-supervised multi-organ segmentation,

    Y . Zhou, Z. Li, S. Bai, C. Wang, X. Chen, M. Han, E. Fishman, and A. L. Yuille, “Prior-aware neural network for partially-supervised multi-organ segmentation,” inICCV, 2019, pp. 10 672–10 681

    2019

  9. [17]

    Multi-structure segmentation from partially labeled datasets. application to body com- position measurements on ct scans,

    G. Gonz ´alez, G. R. Washko, and R. San Jos ´e Est ´epar, “Multi-structure segmentation from partially labeled datasets. application to body com- position measurements on ct scans,” inInternational Workshop on Reconstruction and Analysis of Moving Body Organs. Springer, 2018, ...

    2018

  10. [18]

    Multi-organ segmentation via co-training weight-averaged models from few-organ datasets,

    R. Huang, Y . Zheng, Z. Hu, S. Zhang, and H. Li, “Multi-organ segmentation via co-training weight-averaged models from few-organ datasets,” inMICCAI. Springer, 2020, pp. 146–155

    2020

  11. [19]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” inICCV, 2021, pp. 9650–9660

    2021

  12. [20]

    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. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”ICLR, 2020

    2020

  13. [21]

    Unsuper- vised object localization in the era of self-supervised vits: A survey,

    O. Simeoni, E. Zablocki, S. Gidaris, G. Puy, and P. Perez, “Unsuper- vised object localization in the era of self-supervised vits: A survey,” International Journal of Computer Vision, vol. 133, no. 2, pp. 781–808, 2025

    2025

  14. [22]

    Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut,

    Y . Wang, X. Shen, Y . Yuan, Y . Du, M. Li, S. X. Hu, J. L. Crowley, and D. Vaufreydaz, “Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 12, pp. ...

    2023

  15. [23]

    Self-supervised transformers for unsupervised object discovery using normalized cut,

    Y . Wang, X. Shen, S. X. Hu, Y . Yuan, J. L. Crowley, and D. Vaufreydaz, “Self-supervised transformers for unsupervised object discovery using normalized cut,” inCVPR, 2022, pp. 14 543–14 553

    2022

  16. [24]

    Cut and learn for unsupervised object detection and instance segmentation,

    X. Wang, R. Girdhar, S. X. Yu, and I. Misra, “Cut and learn for unsupervised object detection and instance segmentation,” inCVPR, 2023, pp. 3124–3134

    2023

  17. [25]

    Towards general purpose vision foundation models for medical image analysis: An experimental study of dinov2 on radiology benchmarks,

    M. Baharoon, W. Qureshi, J. Ouyang, Y . Xu, A. Aljouie, and W. Peng, “Towards general purpose vision foundation models for medical image analysis: An experimental study of dinov2 on radiology benchmarks,”arXiv preprint arXiv:2312.02366, 2023. [Online]. Available: https://arxiv...

    2023 arXiv

  18. [26]

    Exploring scalable medical image encoders beyond text supervision,

    F. P ´erez-Garc´ıa, H. Sharma, S. Bond-Taylor, K. Bouzid, V . Salvatelli, M. Ilse, S. Bannur, D. C. Castro, A. Schwaighofer, M. P. Lungrenet al., “Exploring scalable medical image encoders beyond text supervision,” Nature Machine Intelligence, pp. 1–12, 2025

    2025

  19. [27]

    Xception: Deep learning with depthwise separable convolu- tions,

    F. Chollet, “Xception: Deep learning with depthwise separable convolu- tions,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1251–1258

    2017

  20. [28]

    Multi-modal learning from unpaired images: Application to multi-organ segmentation in ct and mri,

    V . V . Valindria, N. Pawlowski, M. Rajchl, I. Lavdas, E. O. Aboagye, A. G. Rockall, D. Rueckert, and B. Glocker, “Multi-modal learning from unpaired images: Application to multi-organ segmentation in ct and mri,” inWACV. IEEE, 2018, pp. 547–556

    2018

  21. [29]

    The liver tumor segmentation benchmark (lits),

    P. Bilic, P. Christ, H. B. Li, E. V orontsov, A. Ben-Cohen, G. Kaissis, A. Szeskin, C. Jacobs, G. E. H. Mamani, G. Chartrandet al., “The liver tumor segmentation benchmark (lits),”Medical image analysis, vol. 84, p. 102680, 2023

    2023

  22. [30]

    An international challenge to use artificial intelligence to define the state- of-the-art in kidney and kidney tumor segmentation in ct imaging

    N. Heller, S. McSweeney, M. T. Peterson, S. Peterson, J. Rickman, B. Stai, R. Tejpaul, M. Oestreich, P. Blake, J. Rosenberget al., “An international challenge to use artificial intelligence to define the state- of-the-art in kidney and kidney tumor segmentation in ct imaging.” 2020

    2020

  23. [31]

    Abdomenct-1k: Is abdominal organ segmen- tation a solved problem?

    J. Ma, Y . Zhang, S. Gu, C. Zhu, C. Ge, Y . Zhang, X. An, C. Wang, Q. Wang, X. Liuet al., “Abdomenct-1k: Is abdominal organ segmen- tation a solved problem?”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 6695–6714, 2021

    2021

  24. [32]

    Ct-org, a new dataset for multiple organ segmentation in computed tomography,

    B. Rister, D. Yi, K. Shivakumar, T. Nobashi, and D. L. Rubin, “Ct-org, a new dataset for multiple organ segmentation in computed tomography,” Scientific Data, vol. 7, no. 1, p. 381, 2020

    2020

  25. [33]

    Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation,

    H. R. Roth, L. Lu, A. Farag, H.-C. Shin, J. Liu, E. B. Turkbey, and R. M. Summers, “Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation,” inMICCAI. Springer, 2015, pp. 556–564

    2015

  26. [34]

    Unleashing the strengths of unlabeled data in pan-cancer abdominal organ quantification: the flare22 challenge,

    J. Ma, Y . Zhang, S. Gu, C. Ge, S. Ma, A. Young, C. Zhu, K. Meng, X. Yang, Z. Huanget al., “Unleashing the strengths of unlabeled data in pan-cancer abdominal organ quantification: the flare22 challenge,”arXiv preprint arXiv:2308.05862, 2023

    2023 arXiv

  27. [35]

    Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation,

    Y . Ji, H. Bai, C. Ge, J. Yang, Y . Zhu, R. Zhang, Z. Li, L. Zhanng, W. Ma, X. Wanet al., “Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation,”Advances in neural information processing systems, vol. 35, pp. 36 722–36 732, 2022

    2022

  28. [36]

    Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge,

    J. Igelsias, M. Styner, T. Langerak, B. Landman, Z. Xu, and A. Klein, “Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge,” inProc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, 2015

    2015

  29. [37]

    Word: A large scale dataset, bench- mark and clinical applicable study for abdominal organ segmentation from ct image,

    X. Luo, W. Liao, J. Xiao, J. Chen, T. Song, X. Zhang, K. Li, D. N. Metaxas, G. Wang, and S. Zhang, “Word: A large scale dataset, bench- mark and clinical applicable study for abdominal organ segmentation from ct image,”Medical Image Analysis, vol. 82, p. 102642, 2022

    2022

  30. [38]

    The medical segmentation decathlon,

    M. Antonelli, A. Reinke, S. Bakas, K. Farahani, A. Kopp-Schneider, B. A. Landman, G. Litjens, B. Menze, O. Ronneberger, R. M. Summers, and B. van Ginneken, “The medical segmentation decathlon,”Nature Communications, vol. 13, no. 1, p. 4128, 2022

    2022

  31. [39]

    Learn2reg: comprehensive multi-task medical image registration challenge, dataset and evaluation in the era of deep learning,

    A. Hering, L. Hansen, T. C. Mok, A. C. Chung, H. Siebert, S. H ¨ager et al., “Learn2reg: comprehensive multi-task medical image registration challenge, dataset and evaluation in the era of deep learning,”IEEE Transactions on Medical Imaging, vol. 42, no. 3, pp. 697–712, 2022

    2022

  32. [40]

    3d image recon- struction for comparison of algorithm database,

    L. Soler, A. Hostettler, V . Agnus, A. Charnoz, J.-B. Fasquel, J. Moreau, A.-B. Osswald, M. Bouhadjar, and J. Marescaux, “3d image recon- struction for comparison of algorithm database,” https://www.ircad.fr/ research/data-sets/liver-segmentation-3d-ircadb-01, 2010

    2010

  33. [41]

    Unet++: A nested u-net architecture for medical image segmentation,

    Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, and J. Liang, “Unet++: A nested u-net architecture for medical image segmentation,” inDLMIA. Springer International Publishing, 2018, pp. 3–11

    2018

  34. [42]

    Transunet: Transformers make strong encoders for medical image segmentation,

    J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, and Y . Zhou, “Transunet: Transformers make strong encoders for medical image segmentation,”arXiv preprint arXiv:2102.04306, 2021

    2021 arXiv

  35. [43]

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

    A. Hatamizadeh, V . Nath, Y . Tang, D. Yang, H. R. Roth, and D. Xu, “Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,” inInternational MICCAI BrainLesion Workshop. Springer International Publishing, 2021, pp. 272–284, cham

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.