Pith. sign in

REVIEW 4 major objections 5 minor 63 references

A concept-driven foundation model, trained on ultrasound-specific image–mask–concept triplets, segments 13 anatomical categories from text prompts alone, outperforming existing text- and concept-driven medical segmentation models.

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 11:44 UTC pith:K4CB2SBZ

load-bearing objection Substantial ultrasound-specific SAM3 adaptation with strong average gains, but the 'consistently outperforms' claim is contradicted by the paper's own tables; worth a serious referee. the 4 major comments →

arxiv 2607.29200 v1 pith:K4CB2SBZ submitted 2026-07-31 cs.CV

UltraSAM3: A Concept-Driven Foundation Model for Universal Ultrasound Image Segmentation

Bo Xu , Quanhao Zhu , Rui Lin , Boling Zhu , Chenyuan Wang , Hongfei Lin , Feng Xia , Chenhua Ji This is my paper
classification cs.CV
keywords ultrasound segmentationconcept-driven segmentationfoundation modeltext promptsSAM3multi-organ segmentationsemantic groundinginstruction-guided agent
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.

UltraSAM3 adapts the concept-prompted segmentation architecture SAM3 to ultrasound by fine-tuning it on a large corpus of 37 public ultrasound datasets covering 13 organs, where every training example is an image–mask–concept triplet with a short medical phrase as the concept. The paper argues that this ultrasound-specific adaptation gives the model reliable semantic grounding in noisy, low-contrast ultrasound images, so that a user can segment a target simply by naming it—no clicks or bounding boxes. A companion instruction-guided agent converts complex natural-language requests into concise concept prompts, improving robustness when users ask full questions. If correct, this points toward a single interactive model for diverse ultrasound reading tasks that currently require separate task-specific pipelines.

Core claim

The paper claims that full-parameter fine-tuning of SAM3 on ultrasound image–mask–concept triplets yields a universal ultrasound segmentation model that can be prompted by short medical concepts and outperforms representative concept- and text-driven biomedical segmentation baselines on multi-organ benchmarks, external datasets, and visual-prompt-enhanced settings. The average Dice over 13 representative categories reaches 0.7144 (IoU 0.6342), with consistent gains across most organs; the largest margins appear in thyroid, muscle, prostate, and cardiac structures. The instruction-guided agent, built on a large multimodal language model, further raises average Dice by 0.161 over directly feed

What carries the argument

The core machinery is the concept-prompted segmentation interface of SAM3, adapted through full-parameter fine-tuning on ultrasound-specific triplets (image, mask, concept). The image encoder, text encoder, detector, tracker, and segmentation modules are all updated, so the model learns to align noisy ultrasound appearances with short anatomical or lesion phrases. An instruction-guided agent sits on top: it selects the most relevant organ-level concept from a candidate pool and rewrites a complex user query into a concise prompt such as 'breast lesion' or 'thyroid nodule', which is then fed to UltraSAM3 for mask prediction.

Load-bearing premise

The reported gains presuppose that converting ground-truth labels from 37 heterogeneous datasets into standardized short concept phrases yields a coherent, unambiguous label space, and that short noun phrases give the model enough semantic grounding in noisy ultrasound images.

What would settle it

A controlled experiment that swaps or shuffles concept prompts across organs (e.g., prompting 'thyroid nodule' on a liver image) and shows that segmentation accuracy does not degrade would falsify the claim of true concept grounding; conversely, demonstrating a low-contrast target that no short phrase can disambiguate would falsify the universality claim.

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

If this is right

  • If confirmed, ultrasound segmentation could move from a collection of task-specific models to one text-promptable system that handles organs, lesions, and anatomical structures in a single pass.
  • The framework suggests that concept-driven prompting can replace visual prompts (boxes, points) for many common ultrasound tasks, reducing the need for expert localization before segmentation.
  • The agent-based instruction parsing implies that natural-language clinical queries can be made compatible with segmentation models without retraining them, by decomposing understanding from pixel prediction.
  • The reported generalization to external datasets indicates the learned ultrasound-aware representations may transfer across scanners and acquisition protocols better than general medical segmentation models.
  • Because both text-only and visual-prompt-enhanced performance improve, the ultrasound-specific adaptation appears to strengthen the underlying visual-semantic alignment rather than merely adding a text shortcut.

Where Pith is reading between the lines

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

  • A direct test of whether the gains reflect true concept grounding would be to corrupt concepts (e.g., swap 'thyroid nodule' with 'liver lesion') and measure performance drop; if the model resists such swaps, it may be relying on image priors more than the textual concept.
  • The instruction-guided agent's +0.161 Dice gain is measured on instructions generated by a proprietary commercial language model; the gain might shrink or vary if the parser were tested on human clinical queries or a different language model.
  • The manual concept normalization that merges diverse dataset labels into phrases like 'breast lesion' may mask label-ambiguity problems; a useful extension would assess segmentation quality when a single short phrase covers visually heterogeneous subcategories (e.g., benign vs. malignant tumors).
  • If concept-driven segmentation scales, one could imagine open-vocabulary ultrasound settings where a model like UltraSAM3 is prompted with free-text findings from a radiology report, enabling automated measurement generation.

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

4 major / 5 minor

Summary. The paper proposes UltraSAM3, a concept-driven ultrasound segmentation model obtained by full fine-tuning of SAM3 on 37 public ultrasound datasets (13 anatomical categories) using image–mask–concept triplets. It also introduces an instruction-guided agent based on Gemini 3.1 Pro that rewrites complex user queries into concise concept prompts. The authors report large IoU/Dice gains over UniBiomed, BiomedParse, SAM3, and Medical SAM3 on a 13-dataset representative benchmark, on external datasets, and under visual-prompt-enhanced settings, and claim that the agent improves robustness for complex instructions.

Significance. If the claims are validated, this is a substantial contribution: the first ultrasound-specific concept-driven segmentation foundation model, trained on a large corpus (112,634 images / 171,693 masks from 37 datasets, 13 categories), with external generalization tests and full per-dataset tables. The external evaluation (Table 2) shows UltraSAM3 winning clearly on all four held-out datasets, and code is promised. These are real strengths. However, the flagship 'consistently outperforms' claim is not supported by the paper's own full results, and the agent experiment needs de-circularization.

major comments (4)
  1. [Abstract; §4.3/C, Tables 4–6] The claim that UltraSAM3 'consistently outperforms' is contradicted by the paper's own full tables. Negative IoU/Dice deltas occur on BUID (IoU 0.8119 vs 0.8162; Dice 0.8809 vs 0.8924), BUSI (IoU 0.5912 vs 0.7950; Dice 0.6570 vs 0.8665), CAMUS (IoU 0.3353 vs 0.5638; Dice 0.4989 vs 0.6145), CardiacUDC (IoU 0.2137 vs 0.2520; Dice 0.3322 vs 0.3708), EchoCP (IoU 0.2791 vs 0.3561; Dice 0.4312 vs 0.5104), and Unity (Dice 0.6186 vs 0.6890). The average gains are driven by thyroid, muscle, and prostate; cardiac and several breast datasets show the reverse. No error bars, multiple seeds, or significance tests are reported, so even the average cannot be separated from run-to-run variability. Please rephrase the headline claim and report per-dataset win/loss counts and confidence intervals.
  2. [§4.6, Fig. 4, Appendix B] The instruction-guided agent experiment is self-referential. The complex instructions are generated by Gemini 3.1 Pro (Sec 4.6), and the same Gemini 3.1 Pro is used as the parsing agent (Fig. 2, Sec 3.4). Because the generator and parser are the same model, the reported +0.161 average Dice gain may reflect prompt self-compatibility rather than robustness to realistic user queries. Please evaluate on human-written instructions or a different generator/parser, and report instruction samples and parsing failures.
  3. [§3.3, §4.1, Appendix B] The manual normalization of raw labels into concept phrases is not documented. The mapping from each of the 37 datasets to its concept set is not provided, and no measure of concept ambiguity or inter-annotator agreement is given. Since the text concept is the only conditioning signal, inconsistent normalization could mean the gains reflect label alignment rather than a generalizable ultrasound representation. Please release the concept mapping and analyze per-concept and per-dataset performance.
  4. [§4.2, Tables 1, 4–6] The text-prompt protocol for the baselines (UniBiomed, BiomedParse, SAM3, Medical SAM3) is not specified. To make the comparison fair, all models must receive the same normalized concept phrase and comparable decoding settings. Without documenting the prompts and any prompt templates used for each baseline, the reported deltas cannot be fully verified. Please provide this information.
minor comments (5)
  1. [Fig. 3] Caption says four radar plots (mAP, AP50, IoU, Dice), but only IoU and Dice panels are shown. Add the missing panels or correct the caption.
  2. [Throughout] Naming: 'US-SAM3' and 'UltraSAM3' are used interchangeably; standardize to one name.
  3. [Eqs. (3), (18)] K and area thresholds τmin/τmax are never specified; state their values and provide an ablation if they affect results.
  4. [Fig. 5, §4.5] Visual-prompt results are presented only graphically without a numeric table or error bars; include numbers for reproducibility.
  5. [Fig. 4, §4.6] Only 50 test images per dataset are sampled for the agent study; this is small for several datasets and the caption should report confidence intervals or variance.

Circularity Check

1 steps flagged

No significant circularity in the segmentation claims; one auxiliary agent evaluation is mildly self-referential.

specific steps
  1. other [Section 4.6 (Instruction-guided Agent for Complex User Queries) and Appendix B (Agent Prompt Template)]
    "For each dataset, we randomly sample 50 test images and use Gemini-3.1-Pro to generate complex but unambiguous user instructions according to the ground-truth target category. ... You are a lightweight front-end agent for SAM3 ultrasound segmentation. ... If the category list contains the target, choose exactly one option from it."

    In Eq. (5)/(7)-(8), the agent gphi produces the concept prompt q* = gphi(I, u, Co). The evaluation of the agent's benefit constructs the complex instruction u with the same Gemini 3.1 Pro model that later parses u, and Co contains the ground-truth category. Thus the parsed prompt is the very category the instruction was synthesized from, so the reported +0.161 Dice improvement over the direct-instruction baseline is partly ensured by the evaluation design rather than by independent human-query testing. This self-reference is confined to the secondary usability/agent experiment and does not enter the training or external-dataset evaluation of the segmentation model, so it is not load-bearing for the central claim.

full rationale

The central derivation is standard supervised concept-conditioned segmentation: image-mask-concept triplets are built from public dataset labels, SAM3 is fine-tuned, and held-out external datasets (BrEast, CCAUI, 105US_tumor, KFGNet) provide an independent generalization check. There is no fitted parameter renamed as a prediction, no load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in via citation. The abstract's 'consistently outperforms' wording is not supported by the paper's own full tables (e.g., BUID IoU 0.8119 vs BiomedParse 0.8162; BUSI IoU 0.5912 vs 0.7950; CAMUS IoU 0.3353 vs 0.5638), but an overstated headline is a correctness/claims-consistency issue rather than a circularity issue. The only self-referential element is the agent experiment, where Gemini 3.1 Pro both generates and parses the complex instructions; this is a minor methodological weakness in a secondary evaluation and does not compromise the independent segmentation results. Score 1 reflects that the core model evaluation is self-contained and externally grounded.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central empirical claims rest on SAM3's transferability, label harmonization across 37 datasets, and the closed-source Gemini agent. These are domain assumptions rather than tested hypotheses. The main unstated hand-chosen components are the mask-selection thresholds and candidate mask count; training hyperparameters are absent. No new theoretical entities are introduced.

free parameters (4)
  • Area thresholds τmin, τmax (Eq. 18) = not reported
    Candidate masks are filtered by these thresholds before final selection; values are never given, so reported scores depend on unstated hand-chosen numbers.
  • Candidate mask count K (Eq. 3) = not reported
    Number of mask candidates generated per concept is unspecified and affects the final pick.
  • Training hyperparameters (optimizer, LR, epochs, batch size) = not reported
    Full-parameter fine-tuning protocol in Section 3.3 gives no reproducibility-critical settings.
  • Per-organ concept pools Co (Appendix B) = 13 hand-crafted pools from dataset labels
    The agent selects among organ-level candidate sets; composition of these pools is not systematically specified or validated.
axioms (4)
  • domain assumption SAM3 concept-conditioned segmentation can be transferred to ultrasound by full fine-tuning.
    Method relies on adapting SAM3's concept mechanism to ultrasound (Section 3.3).
  • domain assumption Labels of 37 public datasets can be normalized to 13 shared categories and expressed as short text concepts.
    Training triplets require manual normalization of concepts (Sections 3.3, 4.1); cross-dataset label consistency is assumed, not analyzed.
  • domain assumption Text-only concept prompts are a clinically useful interaction mode.
    Motivation claims clinicians want text queries (Section 1, 3.1); no user study is provided.
  • domain assumption Gemini 3.1 Pro can parse instructions and select concepts reliably.
    Agent design depends on a closed-source LLM (Section 4.6, Appendix B); performance is not reproducible.

pith-pipeline@v1.3.0-daily-deepseek · 22000 in / 14478 out tokens · 141322 ms · 2026-08-03T11:44:53.251244+00:00 · methodology

0 comments
read the original abstract

Ultrasound imaging has become increasingly widespread in clinical practice due to its portability, low cost and real-time capability, making ultrasound image segmentation important. However, ultrasound images differ substantially from CT, MRI, and other medical imaging modalities, as they are often affected by speckle noise, low contrast, acoustic shadows and ambiguous boundaries. Existing ultrasound segmentation methods are still mainly limited to task-specific models or visual-prompt-based foundation models, which are either tailored to particular tasks or require expert-provided visual prompts, making them inconvenient for flexible clinical use. To address these challenges, we propose UltraSAM3, a concept-driven foundation model for universal ultrasound image segmentation. Unlike conventional models, UltraSAM3 enables text-based target specification by adapting SAM3 to ultrasound-specific image--mask--concept triplets. The model is trained on a large-scale ultrasound segmentation corpus covering 37 public datasets and 13 anatomical categories, allowing it to align ultrasound visual patterns with clinically meaningful concepts across diverse organs and lesions. To further improve usability under realistic clinical interaction, we propose an instruction-guided agent that parses complex natural language queries into concise ultrasound concept prompts for UltraSAM3. Extensive experiments demonstrate that UltraSAM3 consistently outperforms representative concept- and text-driven biomedical segmentation models on multi-organ ultrasound benchmarks, external datasets, and visual-prompt-enhanced settings. Moreover, the agent improves segmentation robustness for complex user instructions. These results indicate that ultrasound-specific concept adaptation is effective for building generalizable and interactive ultrasound segmentation foundation models.

Figures

Figures reproduced from arXiv: 2607.29200 by Boling Zhu, Bo Xu, Chenhua Ji, Chenyuan Wang, Feng Xia, Hongfei Lin, Quanhao Zhu, Rui Lin.

Figure 1
Figure 1. Figure 1: Overview of UltraSAM3 for concept-driven universal ultrasound segmentation. Given ultrasound images and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of UltraSAM3 and the instruction-guided agent. The upper part illustrates the construction [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Organ-wise performance comparison on the full ultrasound corpus. All 37 datasets are grouped into 13 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Dice improvement achieved by the instruction-guided agent over directly feeding complex instructions into [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: IoU comparison under visual-prompt-enhanced segmentation. Each row corresponds to one ultrasound [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of different concept-driven segmentation models on representative ultrasound cases. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Organ-level data distribution of the 37 ultrasound datasets used in this work. [PITH_FULL_IMAGE:figures/full_fig_p016_7.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

63 extracted references · 7 linked inside Pith

  1. [1]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015

    2015

  2. [2]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation.Nature methods, 18(2):203–211, 2021

    Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation.Nature methods, 18(2):203–211, 2021

    2021

  3. [3]

    Transunet: Transformers make strong encoders for medical image segmentation.arXiv preprint arXiv:2102.04306, 2021

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation.arXiv preprint arXiv:2102.04306, 2021

    Pith/arXiv arXiv 2021

  4. [4]

    Unetr: Transformers for 3d medical image segmentation

    Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 574–584, 2022

    2022

  5. [5]

    U-mamba: Enhancing long-range dependency for biomedical image segmentation

    Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Vm-unet: Vision mamba unet for medical image segmentation

    Jiacheng Ruan, Jincheng Li, and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. ACM Transactions on Multimedia Computing, Communications and Applications, 2024

    2024

  7. [7]

    Swin-umamba: Mamba-based unet with imagenet-based pretraining

    Jiarun Liu, Hao Yang, Hong-Yu Zhou, Yan Xi, Lequan Yu, Cheng Li, Yong Liang, Guangming Shi, Yizhou Yu, Shaoting Zhang, et al. Swin-umamba: Mamba-based unet with imagenet-based pretraining. InInternational conference on medical image computing and computer-assisted intervention, pages 615–625. Springer, 2024

    2024

  8. [8]

    Log-vmamba: local-global vision mamba for medical image segmentation

    Trung Dinh Quoc Dang, Huy Hoang Nguyen, and Aleksei Tiulpin. Log-vmamba: local-global vision mamba for medical image segmentation. InProceedings of the Asian Conference on Computer Vision, pages 548–565, 2024

    2024

  9. [9]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. InProceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    2023

  10. [10]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Segment anything in medical images.Nature communications, 15(1):654, 2024

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images.Nature communications, 15(1):654, 2024

    2024

  12. [12]

    Medsam2: Segment anything in 3d medical images and videos.arXiv preprint arXiv:2504.03600, 2025

    Jun Ma, Zongxin Yang, Sumin Kim, Bihui Chen, Mohammed Baharoon, Adibvafa Fallahpour, Reza Asakereh, Hongwei Lyu, and Bo Wang. Medsam2: Segment anything in 3d medical images and videos.arXiv preprint arXiv:2504.03600, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Ultrasam: a foundation model for ultrasound using large open-access segmentation datasets.International Journal of Computer Assisted Radiology and Surgery, pages 1–10, 2025

    Adrien Meyer, Aditya Murali, Farahdiba Zarin, Didier Mutter, and Nicolas Padoy. Ultrasam: a foundation model for ultrasound using large open-access segmentation datasets.International Journal of Computer Assisted Radiology and Surgery, pages 1–10, 2025

    2025

  14. [14]

    Samusa: Segment anything model 2 for ultrasound annotation

    Baptiste Podvin, Toby Collins, Güinther Saibro, Chiara Innocenzi, Yuchuan Yang, Flavio Milana, Yvonne Keeza, Grace Ufitinema, Florien Ujemurwego, Guido Torzilli, et al. Samusa: Segment anything model 2 for ultrasound annotation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 509–518. Springer, 2025

    2025

  15. [15]

    Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025. 12 Running Title for Header

    Pith/arXiv arXiv 2025

  16. [16]

    Biomedparse: a biomedical foundation model for image parsing of everything everywhere all at once.arXiv preprint arXiv:2405.12971, 2024

    Theodore Zhao, Yu Gu, Jianwei Yang, Naoto Usuyama, Ho Hin Lee, Tristan Naumann, Jianfeng Gao, Angela Crabtree, Jacob Abel, Christine Moung-Wen, et al. Biomedparse: a biomedical foundation model for image parsing of everything everywhere all at once.arXiv preprint arXiv:2405.12971, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Large-vocabulary segmentation for medical images with text prompts.NPJ Digital Medicine, 8(1):566, 2025

    Ziheng Zhao, Yao Zhang, Chaoyi Wu, Xiaoman Zhang, Xiao Zhou, Ya Zhang, Yanfeng Wang, and Weidi Xie. Large-vocabulary segmentation for medical images with text prompts.NPJ Digital Medicine, 8(1):566, 2025

    2025

  18. [18]

    Unibiomed: A universal foundation model for grounded biomedical image interpretation.arXiv preprint arXiv:2504.21336, 2025

    Linshan Wu, Yuxiang Nie, Sunan He, Jiaxin Zhuang, Luyang Luo, Tao Li, Zhuoyao Xie, Dexuan Chen, Yinghua Zhao, Neeraj Mahboobani, et al. Unibiomed: A universal foundation model for grounded biomedical image interpretation.arXiv preprint arXiv:2504.21336, 2025

    arXiv 2025

  19. [19]

    Medsam3: Delving into segment anything with medical concepts.arXiv preprint arXiv:2511.19046, 2025

    Anglin Liu, Rundong Xue, Xu R Cao, Yifan Shen, Yi Lu, Xiang Li, Qianqian Chen, and Jintai Chen. Medsam3: Delving into segment anything with medical concepts.arXiv preprint arXiv:2511.19046, 2025

    arXiv 2025

  20. [20]

    Medical sam3: A foundation model for universal prompt-driven medical image segmentation.arXiv preprint arXiv:2601.10880, 2026

    Chongcong Jiang, Tianxingjian Ding, Chuhan Song, Jiachen Tu, Ziyang Yan, Yihua Shao, Zhenyi Wang, Yuzhang Shang, Tianyu Han, and Yu Tian. Medical sam3: A foundation model for universal prompt-driven medical image segmentation.arXiv preprint arXiv:2601.10880, 2026

    arXiv 2026

  21. [21]

    Kd-eye: Lightweight pupil segmentation for eye tracking on vr headsets via knowledge distillation

    Yanlin Li, Ning Chen, Guangrong Zhao, and Yiran Shen. Kd-eye: Lightweight pupil segmentation for eye tracking on vr headsets via knowledge distillation. InInternational Conference on Wireless Artificial Intelligent Computing Systems and Applications, pages 209–220. Springer, 2024

    2024

  22. [22]

    Echogpt: An interactive cardiac function assessment model for echocardiogram videos

    Bo Xu, Quanhao Zhu, Qingchen Zhang, Mengmeng Wang, Liang Zhao, Hongfei Lin, Jing Ren, and Feng Xia. Echogpt: An interactive cardiac function assessment model for echocardiogram videos. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pages 7858–7866, 2025

    2025

  23. [23]

    Segment anything model 2: an application to 2d and 3d medical images.IEEE Transactions on Biomedical Engineering, 2026

    Haoyu Dong, Hanxue Gu, Yaqian Chen, Jichen Yang, Yuwen Chen, and Maciej A Mazurowski. Segment anything model 2: an application to 2d and 3d medical images.IEEE Transactions on Biomedical Engineering, 2026

    2026

  24. [24]

    AbdomenUS: US Simulation and Segmentation

    José Ignacio Orlando. AbdomenUS: US Simulation and Segmentation. https://www.kaggle.com/datasets/ ignaciorlando/ussimandsegm, 2018

    2018

  25. [25]

    Bus-bra: A breast ultrasound dataset for assessing computer-aided diagnosis systems.Medical physics, 51(4):3110–3123, 2024

    Wilfrido Gómez-Flores, Maria Julia Gregorio-Calas, and Wagner Coelho de Albuquerque Pereira. Bus-bra: A breast ultrasound dataset for assessing computer-aided diagnosis systems.Medical physics, 51(4):3110–3123, 2024

    2024

  26. [26]

    Video-based ai for beat-to-beat assessment of cardiac function.Nature, 580(7802):252–256, 2020

    David Ouyang, Bryan He, Amirata Ghorbani, Neal Yuan, Joseph Ebinger, Curtis P Langlotz, Paul A Heidenreich, Robert A Harrington, David H Liang, Euan A Ashley, et al. Video-based ai for beat-to-beat assessment of cardiac function.Nature, 580(7802):252–256, 2020

    2020

  27. [27]

    Mi-segnet: Mutual information-based us segmentation for unseen domain generalization

    Yuan Bi, Zhongliang Jiang, Ricarda Clarenbach, Reza Ghotbi, Angelos Karlas, and Nassir Navab. Mi-segnet: Mutual information-based us segmentation for unseen domain generalization. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 130–140. Springer, 2023

    2023

  28. [28]

    Pubic symphysis-fetal head segmentation and angle of progression, March 2024

    Bai Jieyun and Ou ZhanHong. Pubic symphysis-fetal head segmentation and angle of progression, March 2024

    2024

  29. [29]

    Ultrasound Normal Kidney Image Dataset

    JeevaWS. Ultrasound Normal Kidney Image Dataset. https://universe.roboflow.com/jeevaws/ ultrasound-normal-kidney-image, 2025

    2025

  30. [30]

    Annotated ultrasound liver images, November 2022

    Xu Yiming, Zheng Bowen, Liu Xiaohong, Wu Tao, Ju Jinxiu, Wang Shijie, Lian Yufan, Zhang Hongjun, Liang Tong, Sang Ye, Jiang Rui, Wang Guangyu, Ren Jie, and Chen Ting. Annotated ultrasound liver images, November 2022

    2022

  31. [31]

    Lung ultrasound covid phantom dataset used for training machine learning model

    James R McLaughlan, Lewis Howell, and Nicola Ingram. Lung ultrasound covid phantom dataset used for training machine learning model. 2024

    2024

  32. [32]

    Deep learning segmentation of transverse musculoskeletal ultrasound images for neuromuscular disease assessment.Computers in biology and medicine, 135:104623, 2021

    Francesco Marzola, Nens Van Alfen, Jonne Doorduin, and Kristen M Meiburger. Deep learning segmentation of transverse musculoskeletal ultrasound images for neuromuscular disease assessment.Computers in biology and medicine, 135:104623, 2021

    2021

  33. [33]

    Mallesnet: A multi-object assistance based network for brachial plexus segmentation in ultrasound images.Medical Image Analysis, 80:102511, 2022

    Yi Ding, IEEE Member, Qiqi Yang, Yiqian Wang, Dajiang Chen, Zhiguang Qin, and Jian Zhang. Mallesnet: A multi-object assistance based network for brachial plexus segmentation in ultrasound images.Medical Image Analysis, 80:102511, 2022

    2022

  34. [34]

    Mmotu: A multi-modality ovarian tumor ultrasound image dataset for unsupervised cross-domain semantic segmentation.arXiv preprint arXiv:2207.06799, 2022

    Qi Zhao, Shuchang Lyu, Wenpei Bai, Linghan Cai, Binghao Liu, Guangliang Cheng, Meijing Wu, Xiubo Sang, Min Yang, and Lijiang Chen. Mmotu: A multi-modality ovarian tumor ultrasound image dataset for unsupervised cross-domain semantic segmentation.arXiv preprint arXiv:2207.06799, 2022

    Pith/arXiv arXiv 2022

  35. [35]

    Micro-ultrasound prostate segmentation dataset.URL: https://doi

    W Shao and W Brisbane. Micro-ultrasound prostate segmentation dataset.URL: https://doi. org/10.5281/zenodo, 10475293, 2024

    doi:10.5281/zenodo 2024

  36. [36]

    Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules.Computers in biology and medicine, 155:106389, 2023

    Haifan Gong, Jiaxin Chen, Guanqi Chen, Haofeng Li, Guanbin Li, and Fei Chen. Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules.Computers in biology and medicine, 155:106389, 2023. 13 Running Title for Header

    2023

  37. [37]

    Luminous database: lumbar multifidus muscle segmentation from ultrasound images.BMC Musculoskeletal Disorders, 21(1):703, 2020

    Clyde J Belasso, Bahareh Behboodi, Habib Benali, Mathieu Boily, Hassan Rivaz, and Maryse Fortin. Luminous database: lumbar multifidus muscle segmentation from ultrasound images.BMC Musculoskeletal Disorders, 21(1):703, 2020

    2020

  38. [38]

    FALLMUD: FAscicle Lower Leg Muscle Ultrasound Dataset

    Rebecca Cunningham and Neil Cronin. FALLMUD: FAscicle Lower Leg Muscle Ultrasound Dataset. https: //kalisteo.cea.fr/index.php/fallmud/, 2020

    2020

  39. [39]

    Deep learning for segmentation using an open large-scale dataset in 2d echocardiography.IEEE transactions on medical imaging, 38(9):2198– 2210, 2019

    Sarah Leclerc, Erik Smistad, Joao Pedrosa, Andreas Østvik, Frederic Cervenansky, Florian Espinosa, Torvald Espeland, Erik Andreas Rye Berg, Pierre-Marc Jodoin, Thomas Grenier, et al. Deep learning for segmentation using an open large-scale dataset in 2d echocardiography.IEEE transactions on medical imaging, 38(9):2198– 2210, 2019

    2019

  40. [40]

    Unity Imaging Echocardiography Datasets

    Unity Imaging Collaborative. Unity Imaging Echocardiography Datasets. https://data.unityimaging.net/, 2024

    2024

  41. [41]

    Echocp: An echocardiography dataset in contrast transthoracic echocardiography for patent foramen ovale diagnosis

    Tianchen Wang, Zhihe Li, Meiping Huang, Jian Zhuang, Shanshan Bi, Jiawei Zhang, Yiyu Shi, Hongwen Fei, and Xiaowei Xu. Echocp: An echocardiography dataset in contrast transthoracic echocardiography for patent foramen ovale diagnosis. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 506–515. Springer, 2021

    2021

  42. [42]

    Charitha D Reddy, Leo Lopez, David Ouyang, James Y Zou, and Bryan He. Video-based deep learning for automated assessment of left ventricular ejection fraction in pediatric patients.Journal of the American Society of Echocardiography, 36(5):482–489, 2023

    2023

  43. [43]

    Graphecho: Graph-driven unsuper- vised domain adaptation for echocardiogram video segmentation

    Jiewen Yang, Xinpeng Ding, Ziyang Zheng, Xiaowei Xu, and Xiaomeng Li. Graphecho: Graph-driven unsuper- vised domain adaptation for echocardiogram video segmentation. InProceedings of the IEEE/CVF international conference on computer vision, pages 11878–11887, 2023

    2023

  44. [44]

    Mr to ultrasound registration for prostate challenge-dataset

    Zachary Baum, Shaheer Saeed, Zhe Min, Yipeng Hu, and Dean Barratt. Mr to ultrasound registration for prostate challenge-dataset. InMedical Image Computing and Computer Assisted Intervention–MICCAI, volume 2023, 2023

    2023

  45. [45]

    Thyroid Ultrasound Cine-clip, 2024

    Stanford Center for Artificial Intelligence in Medicine and Imaging. Thyroid Ultrasound Cine-clip, 2024

    2024

  46. [46]

    Multi-task learning for thyroid nodule segmentation with thyroid region prior

    Haifan Gong, Guanqi Chen, Ranran Wang, Xiang Xie, Mingzhi Mao, Yizhou Yu, Fei Chen, and Guanbin Li. Multi-task learning for thyroid nodule segmentation with thyroid region prior. In2021 IEEE 18th international symposium on biomedical imaging (ISBI), pages 257–261. IEEE, 2021

    2021

  47. [47]

    Tracked 3d ultrasound and deep neural network-based thyroid segmentation reduce interobserver variability in thyroid volumetry.Plos one, 17(7):e0268550, 2022

    Markus Krönke, Christine Eilers, Desislava Dimova, Melanie Köhler, Gabriel Buschner, Lilit Schweiger, Lemonia Konstantinidou, Marcus Makowski, James Nagarajah, Nassir Navab, et al. Tracked 3d ultrasound and deep neural network-based thyroid segmentation reduce interobserver variability in thyroid volumetry.Plos one, 17(7):e0268550, 2022

    2022

  48. [48]

    An open access thyroid ultrasound image database

    Lina Pedraza, Carlos Vargas, Fabián Narváez, Oscar Durán, Emma Muñoz, and Eduardo Romero. An open access thyroid ultrasound image database. In10th International symposium on medical information processing and analysis, volume 9287, pages 188–193. SPIE, 2015

    2015

  49. [49]

    Fast and accurate u-net model for fetal ultrasound image segmentation.Ultrasonic imaging, 44(1):25–38, 2022

    Vahid Ashkani Chenarlogh, Mostafa Ghelich Oghli, Ali Shabanzadeh, Nasim Sirjani, Ardavan Akhavan, Isaac Shiri, Hossein Arabi, Morteza Sanei Taheri, and Mohammad Kazem Tarzamni. Fast and accurate u-net model for fetal ultrasound image segmentation.Ultrasonic imaging, 44(1):25–38, 2022

    2022

  50. [50]

    Acouslic-ai challenge report: Fetal abdominal circumference measurement on blind-sweep ultrasound data from low-income countries.Medical image analysis, 105:103640, 2025

    M Sofia Sappia, Chris L de Korte, Bram van Ginneken, Dean Ninalga, Satoshi Kondo, Satoshi Kasai, Kousuke Hirasawa, Tanya Akumu, Carlos Martín-Isla, Karim Lekadir, et al. Acouslic-ai challenge report: Fetal abdominal circumference measurement on blind-sweep ultrasound data from low-income countries.Medical image analysis, 105:103640, 2025

    2025

  51. [51]

    Fetal abdominal structures segmentation dataset using ultrasonic images.Mendeley Data, 1:1, 2023

    Karine Souza Da Correggio, R Noya Galluzzo, Luıs Otávio Santos, F Soares Muylaert Barroso, T Zimmer- mann Loureiro Chaves, A Sherlley Casimiro Onofre, and Aldo von Wangenheim. Fetal abdominal structures segmentation dataset using ultrasonic images.Mendeley Data, 1:1, 2023

    2023

  52. [52]

    Focus: four-chamber ultrasound image dataset for fetal cardiac biometric measurement.Zenodo10, 5281, 2025

    Wu Songxiong, Zhang Hongyuan, Ye Tingting, X Haoyu, Z Ping, S Qingjun, W Panying, H Bingsheng, D Lei, and W Guangyao. Focus: four-chamber ultrasound image dataset for fetal cardiac biometric measurement.Zenodo10, 5281, 2025

    2025

  53. [53]

    Automated measurement of fetal head circumference using 2d ultrasound images.PloS one, 13(8):e0200412, 2018

    Thomas LA van den Heuvel, Dagmar de Bruijn, Chris L de Korte, and Bram van Ginneken. Automated measurement of fetal head circumference using 2d ultrasound images.PloS one, 13(8):e0200412, 2018

    2018

  54. [54]

    Automated breast ultrasound lesions detection using convolutional neural networks.IEEE journal of biomedical and health informatics, 22(4):1218–1226, 2017

    Moi Hoon Yap, Gerard Pons, Joan Marti, Sergi Ganau, Melcior Sentis, Reyer Zwiggelaar, Adrian K Davison, and Robert Marti. Automated breast ultrasound lesions detection using convolutional neural networks.IEEE journal of biomedical and health informatics, 22(4):1218–1226, 2017. 14 Running Title for Header

    2017

  55. [55]

    Dataset of breast ultrasound images

    Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020

    2020

  56. [56]

    Memory-efficient transformer network with feature fusion for breast tumor segmentation and classification task.Engineering Applications of Artificial Intelligence, 127:107292, 2024

    Ahmed Iqbal and Muhammad Sharif. Memory-efficient transformer network with feature fusion for breast tumor segmentation and classification task.Engineering Applications of Artificial Intelligence, 127:107292, 2024

    2024

  57. [57]

    An open-access breast lesion ultrasound image database: Applicable in artificial intelligence studies.Computers in Biology and Medicine, 152:106438, 2023

    Ali Abbasian Ardakani, Afshin Mohammadi, Mohammad Mirza-Aghazadeh-Attari, and U Rajendra Acharya. An open-access breast lesion ultrasound image database: Applicable in artificial intelligence studies.Computers in Biology and Medicine, 152:106438, 2023

    2023

  58. [58]

    Segmentation and recognition of breast ultrasound images based on an expanded u-net.Plos one, 16(6):e0253202, 2021

    Yanjun Guo, Xingguang Duan, Chengyi Wang, and Huiqin Guo. Segmentation and recognition of breast ultrasound images based on an expanded u-net.Plos one, 16(6):e0253202, 2021

    2021

  59. [59]

    The open kidney ultrasound data set

    Rohit Singla, Cailin Ringstrom, Grace Hu, Victoria Lessoway, John Reid, Christopher Nguan, and Robert Rohling. The open kidney ultrasound data set. InInternational Workshop on Advances in Simplifying Medical Ultrasound, pages 155–164. Springer Nature Switzerland, 2023

    2023

  60. [60]

    Curated benchmark dataset for ultrasound based breast lesion analysis.Scientific Data, 11(1):148, 2024

    Anna Pawłowska, Anna ´Cwierz-Pie´nkowska, Agnieszka Domalik, Dominika Jagu ´s, Piotr Kasprzak, Rafał Matkowski, Łukasz Fura, Andrzej Nowicki, and Norbert ˙Zołek. Curated benchmark dataset for ultrasound based breast lesion analysis.Scientific Data, 11(1):148, 2024

    2024

  61. [61]

    Common Carotid Artery Ultrasound Images, 2022

    Agata Momot. Common Carotid Artery Ultrasound Images, 2022. Accessed: 2026-05-19

    2022

  62. [62]

    Algorithm guided outlining of 105 pancreatic cancer liver metastases in ultrasound.Scientific Reports, 7(1):12779, 2017

    Alexander Hann, Lucas Bettac, Mark M Haenle, Tilmann Graeter, Andreas W Berger, Jens Dreyhaupt, Dieter Schmalstieg, Wolfram G Zoller, and Jan Egger. Algorithm guided outlining of 105 pancreatic cancer liver metastases in ultrasound.Scientific Reports, 7(1):12779, 2017

    2017

  63. [63]

    Please segment the suspicious tumor region in this breast ultrasound image

    Yuchen Wang, Zhongyu Li, Xiangxiang Cui, Liangliang Zhang, Xiang Luo, Meng Yang, and Shi Chang. Key- frame guided network for thyroid nodule recognition using ultrasound videos. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 238–247. Springer, 2022. A Dataset Statistics Table 3 provides detailed statistics ...

    arXiv 2022