Pith. sign in

REVIEW 4 major objections 5 minor 57 references

XBusNet: Text-Guided Breast Ultrasound Segmentation via Multimodal Vision-Language Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A dual-prompt, dual-branch network with automatically assembled text cues reaches top segmentation scores on a breast ultrasound benchmark, with the largest gains on small lesions.

desk verdict A genuinely new architecture, but the test-time two-pass location self-conditioning in Sec. 2.7 inflates the reported gains and is not ablated, so the SOTA claim can't be taken at face value. read the letter →

arxiv 2509.07213 v1 pith:LICAX5BS submitted 2025-09-08 cs.CV cs.AI

Raja Mallina , Bryar Shareef This is my paper
classification cs.CVcs.AI
keywords breastultrasoundsegmentationvision-languagemodelpromptlearningBI-RADSU-NetCLIPsmalllesion
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 proposes XBusNet, a segmentation network for breast ultrasound images that combines two text prompts with two network branches: a global CLIP-ViT branch conditioned on the lesion's size and rough location, and a local U-Net-style branch modulated by clinical descriptors such as shape, margin, and BI-RADS category. The authors argue that this dual design lets text cues steer the model toward the lesion while the local branch preserves boundary detail, and they report that it beats six baselines on the BLU dataset (mean Dice 0.8765, IoU 0.8149), with the largest relative gains on small lesions. The claim matters because prompts are assembled automatically from structured metadata, so the method promises to inject clinical knowledge into segmentation without manual clicking or free-text annotation.

What carries the argument

The load-bearing mechanism is Semantic Feature Adjustment (SFA): a small projection network that converts a prompt embedding into channel-wise scale (γ) and shift (β) parameters, applied to feature maps as an affine transform γ⊙F+β. SFA is inserted into both branches — once in the global branch using the Global Feature Context Prompt (GFCP, verbalizing size and centroid) and at multiple decoder stages of the local branch using the Local Feature Prompt (LFP, verbalizing shape, margin, and BI-RADS). The modulated global and local feature maps are concatenated and refined by two residual blocks before a 1×1 convolution produces the segmentation logits. This organization is what lets text steer

What would settle it

Re-run the five-fold comparison giving every baseline the same first-pass-then-location protocol, or run XBusNet without the location token; if the gap to the closest baseline shrinks or reverses, the claimed superiority depends on the extra inference pass rather than the architecture.

Watch

Extended reading notes

Core claim

The central claim is that a dual-prompt, dual-branch architecture — a global CLIP ViT branch encoding scene-level cues (lesion size and centroid) and a local ResNet50/U-Net branch modulated by attribute prompts (shape, margin, BI-RADS) — achieves state-of-the-art segmentation on the BLU dataset. The paper reports mean Dice of 0.8765 and IoU of 0.8149 across five folds, outperforming six baselines including U-Net, an anatomy-aware model, a SAM adapter, and a CLIP-based ultrasound segmenter. Size-stratified analysis shows the largest gains on small lesions (0–110 pixel length), and ablations show that removing the global branch, the local branch, or the prompt modulation each degrades performa

Load-bearing premise

The reported gains assume that the location prompt estimated from the model's own first-pass prediction is as reliable as the ground-truth centroid used in training, and that this extra inference pass does not bias the comparison against baselines that run only once.

Editorial extensions

If this is right

  • On the BLU benchmark, the full model reports mean Dice 0.8765 and IoU 0.8149, beating U-Net-style, anatomy-aware, and prompt-guided baselines under the same five-fold protocol.
  • The largest relative improvement appears for small lesions (0–110 px length), where Dice reaches 0.8507, suggesting fewer missed regions and spurious activations precisely where pixel evidence is weakest.
  • Ablations attribute the gain to all three components: disabling the global branch drops Dice to 0.8453, disabling the local branch to 0.8572, and disabling SFA to 0.8600.
  • Because prompts are assembled automatically from structured metadata (size, location, shape, margin, BI-RADS), the reported result supports a pipeline that requires no manual clicks or free-text annotations.

Reading between the lines

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

  • The two-pass inference used for the location prompt is the main confound the paper does not isolate; a cleaner comparison would give the same extra pass to every baseline or replace the self-estimated centroid with a learned localizer, to separate the value of the global cue from the extra computation.
  • If the same prompt pipeline transfers to other ultrasound domains with structured descriptors (e.g., thyroid or liver), the recipe — global location/size text plus attribute text with affine modulation — could become a general pattern for small-target segmentation, but this would need external data validation.
  • The reported small-lesion gains suggest a testable extension in screening settings: measure not only Dice but detection rate and false-positive lesions per image, since Dice alone can miss clinically costly whole-lesion misses.
  • The claim that outputs are consistent with BI-RADS descriptors is only supported qualitatively via Grad-CAM; a quantitative alignment study (e.g., whether predicted shape/margin classes match the prompt) would make the interpretability argument testable.
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 XBusNet, a dual-prompt, dual-branch model for breast ultrasound segmentation. A global CLIP/ViT branch is conditioned on lesion size and location via a Global Feature Context Prompt, while a local ResNet50/U-Net branch is modulated by a Local Feature Prompt built from shape, margin, and BI-RADS descriptors through a channel-wise scaling/shifting mechanism (SFA). Prompts are assembled automatically from structured metadata. The method is evaluated on the BLU dataset (252 images after excluding four multi-lesion images) with five-fold cross-validation, reporting mean Dice 0.8765 and IoU 0.8149, outperforming six baselines, with the largest gains on small lesions. Ablations and Grad-CAM visualizations are included.

Significance. If the reported results hold, the paper makes a useful contribution: it demonstrates a reproducible, metadata-to-prompt pipeline for injecting clinical descriptors into a segmentation model, and the dual-branch design with lightweight feature modulation is a reasonable architectural idea. The authors also provide a clear evaluation protocol, five-fold cross-validation, size-stratified analysis, and component ablations. However, the central SOTA claim is weakened by several load-bearing issues: the model receives privileged ground-truth clinical descriptors at test time, the test-time location prompt is produced by an unablated two-pass self-conditioning procedure not available to baselines, and the LFE ablation is confounded. These concerns must be addressed before the claimed superiority over BUSSAM and CLIP-TNseg can be attributed to the architecture.

major comments (4)
  1. [§2.2, §2.8, Table 1] XBusNet receives privileged ground-truth metadata at test time: the local prompt is built from shape, margin, and BI-RADS descriptors in the dataset CSV, and the global prompt uses the true size category. These are expert annotations of the lesion, not inputs available in a real segmentation setting. Baselines receive no such side information. The reported margin over BUSSAM/CLIP-TNseg may therefore reflect information access rather than architecture. Please report a variant that ablates the local prompt (or all metadata-derived prompts) at test time, and/or compare against baselines given the same clinical side information.
  2. [§2.7, Table 1] At validation/test, the location prompt is produced by a first-pass prediction: threshold at tau=0.30, largest connected component, centroid mapped to a quadrant, then a second full inference pass. No ablation reports Dice/IoU for the first pass vs. the second pass, and no baseline is given an analogous self-conditioning loop. The central comparison is therefore between 'XBusNet + two-pass refinement' and single-pass baselines. Please isolate the two-pass mechanism and include a fair two-pass baseline (e.g., coarse mask, centroid feedback, re-segment) to show the architecture alone drives the gains.
  3. [Table 4] The LFE ablation is confounded: the row with LFE=No also has SFA=No, so the drop from 0.8765 to 0.8572 cannot be attributed to the local branch alone. Similarly, removing GFE also removes the global-branch SFA modulation. Please report a full factorial ablation (LFE on/off × GFE on/off × SFA on/off) so each component is isolated.
  4. [Table 2, §3.1] The size-stratified analysis is a headline result, but the table does not state the units or sample sizes of the length intervals (0–110, 111–250, 250+), and no per-bin confidence intervals or significance tests are provided. With only 252 images, the small-lesion bin may contain few cases; please report counts per bin and statistical significance for the small-lesion gains.
minor comments (5)
  1. [Abstract, §3.1, Table 1] Numbers are inconsistent: abstract reports Dice 0.8765 and IoU 0.8149, Table 1 reports 0.877/0.815, and the text says 0.8766/0.8150. Harmonize.
  2. [§2.7] Specify what 'without spatial text' means for the first pass: is the global location token omitted but size and local prompts still used?
  3. [§2.3] 'Five-fold cross-validation' with an 80/20 split per fold is non-standard; clarify whether the folds are non-overlapping test sets.
  4. [§3.1] Wilcoxon test details are missing: number of paired images, whether p-values were corrected for multiple comparisons, and how the five-fold structure was handled.
  5. [References] References [8] and [21] are duplicates; also update arXiv IDs where available.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: central Dice/IoU result is measured against ground truth; the two-pass location prompt is a self-conditioning evaluation confound rather than a circular derivation, and the paper's self-citations are related-work only.

full rationale

Walked the claimed derivation chain. The central result (Dice 0.8765, IoU 0.8149) is a measured comparison against ground-truth masks under five-fold cross-validation (Section 2.4), not a fitted parameter renamed as a prediction. The architecture combines GFE, LFE, and SFA through explicit equations (2)-(17), with components ablated in Table 4. The one self-referential element is the test-time location prompt: Section 2.7 estimates the location from the model's own first-pass probability map (thresholding at tau=0.30, largest connected component, centroid), then feeds it into the final pass as the global prompt. This is a two-stage functional composition, Y = f(I, centroid(threshold(f(I, no-location)))), not an identity: the final mask is not by construction equal to the first-pass mask, and the location token is only a coarse breast quadrant, not the target mask. Therefore it does not meet the 'prediction equals input by construction' test; it is an evaluation-fairness confound (baselines receive one pass, XBusNet receives two, and no ablation isolates the second pass), which is a correctness risk rather than a circularity. The self-citations (Refs. 7, 8, 21, 23, 24) concern BUSIS, STAN/ESTAN, and a prior multitask network; they are related-work context and are not load-bearing for XBusNet's design. No uniqueness theorem is imported, no ansatz is smuggled in via citation, and the method is not a renamed known result. Overall, no load-bearing step reduces to its own input; the score reflects the minor, non-load-bearing self-citations and the unablated self-conditioning step as a methodological caveat, not as circularity.

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

This is an empirical deep learning paper without a derivation. The main ad hoc choices are thresholds and training hyperparameters, none of which are justified by a theory or sensitivity analysis. The model introduces no new physical entities.

free parameters (4)
  • first_pass_threshold_tau = 0.30
    Threshold for converting the first-pass probability map into a coarse proposal used to estimate test-time lesion centroid (Section 2.7). Chosen by hand, no sensitivity analysis.
  • segmentation_threshold_tau_seg = 0.5
    Binarization threshold for computing Dice/IoU metrics (Section 2.4). Fixed without analysis of threshold sensitivity.
  • size_discretization_quantiles = Training-set quantiles (small/medium/large)
    Lesion size is discretized into three bins using quantiles computed from the training split per fold (Section 2.7). Boundaries are data-derived ad hoc choices.
  • training_schedule = 1000 iterations, batch size 4, lr 1e-4, cosine
    Early stopping via fixed iteration count; no tuning or convergence analysis reported (Section 2.3).
assumptions (4)
  • domain assumption Frozen CLIP text encoder produces semantically meaningful embeddings of short clinical phrases (e.g., 'irregular shape, microlobulated margin, BI-RADS 4').
    The prompts are embedded with a frozen CLIP text encoder in Sections 2.7-2.8; if CLIP does not capture clinical semantics, the prompt conditioning is noise.
  • domain assumption Clinical metadata (shape, margin, BI-RADS) is accurate, complete, and informative for segmentation.
    Prompts are verbalized directly from dataset metadata; incorrect or missing descriptors would produce misleading prompts.
  • domain assumption Centroid of the first-pass probability map is a reliable proxy for the true lesion centroid at test time.
    Train-time location uses ground-truth mask centroid; test-time uses a self-estimated centroid (Section 2.7). If first-pass localization fails, the second pass is mis-conditioned.
  • domain assumption The BLU dataset's five-fold split is representative and the 1000-iteration training regime yields stable models.
    Section 2.3 trains only 1000 iterations per fold; the paper assumes this is sufficient for convergence and that fold-wise results generalize within the dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XBusNet: Text-Guided Breast Ultrasound Segmentation via Multimodal Vision-Language Learning." pith.science (2026). https://pith.science/paper/LICAX5BS

@misc{pith2026250907213,
  author       = {Pith},
  title        = {Pith review of: XBusNet: Text-Guided Breast Ultrasound Segmentation via Multimodal Vision-Language Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LICAX5BS}},
  note         = {Machine review of arXiv:2509.07213}
}
read the original abstract

Background: Precise breast ultrasound (BUS) segmentation supports reliable measurement, quantitative analysis, and downstream classification, yet remains difficult for small or low-contrast lesions with fuzzy margins and speckle noise. Text prompts can add clinical context, but directly applying weakly localized text-image cues (e.g., CAM/CLIP-derived signals) tends to produce coarse, blob-like responses that smear boundaries unless additional mechanisms recover fine edges. Methods: We propose XBusNet, a novel dual-prompt, dual-branch multimodal model that combines image features with clinically grounded text. A global pathway based on a CLIP Vision Transformer encodes whole-image semantics conditioned on lesion size and location, while a local U-Net pathway emphasizes precise boundaries and is modulated by prompts that describe shape, margin, and Breast Imaging Reporting and Data System (BI-RADS) terms. Prompts are assembled automatically from structured metadata, requiring no manual clicks. We evaluate on the Breast Lesions USG (BLU) dataset using five-fold cross-validation. Primary metrics are Dice and Intersection over Union (IoU); we also conduct size-stratified analyses and ablations to assess the roles of the global and local paths and the text-driven modulation. Results: XBusNet achieves state-of-the-art performance on BLU, with mean Dice of 0.8765 and IoU of 0.8149, outperforming six strong baselines. Small lesions show the largest gains, with fewer missed regions and fewer spurious activations. Ablation studies show complementary contributions of global context, local boundary modeling, and prompt-based modulation. Conclusions: A dual-prompt, dual-branch multimodal design that merges global semantics with local precision yields accurate BUS segmentation masks and improves robustness for small, low-contrast lesions.

Figures

Figures reproduced from arXiv: 2509.07213 by the authors.

Figure 1
Figure 1. Overview of the proposed XBusNet architecture, showing the Local Feature Extractor (LFE), [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison of breast ultrasound segmentations. False positives (FP) are shown in blue [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Grad-CAM comparison. 4 Discussion XBusNet showed strong performance on BLU. Across five-folds, the model achieved a mean Dice of 0.8765 and an IoU of 0.8149, with narrow fold-to-fold variation, which suggests stable behavior under fold-wise splits. Compared with UNet variants, anatomy-aware models, and text-guided baselines, XBusNet also produced a lower false negative rate, indicating fewer missed lesion pixels. Th… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 39 canonical work pages

  1. [1]

    Key statistics for breast cancer

    American Cancer Society. Key statistics for breast cancer. https://www.cancer.org/cancer/types/ breast-cancer/about/how-common-is-breast-cancer.html, 2025. Last revised May 5, 2025

    work page 2025

  2. [2]

    Nicholson, Michael Silverstein, John B

    Wendy K. Nicholson, Michael Silverstein, John B. Wong, Michael J. Barry, David Chelmow, Tumaini R. Coker, Emilee M. Davis, Carlos R. Jaén, Marie Krousel-Wood, Stacey A. Lee, Lu Li, Carol M. Mangione, Gurkipal S. Rao, John M. Ruiz, James J. Stevermer, Joel Tsevat, Shari M. Underwood, and Sarah E. Wiehe. Screening for breast cancer: US preventive services t...

    work page arXiv 1918

  3. [3]

    Magnetic resonance imaging of the breast: recommendations from the EUSOMA working group.Eur J Cancer, 46(8):1296–1316, 2010

    Francesco Sardanelli, Carla Boetes, Bettina Borisch, and et al. Magnetic resonance imaging of the breast: recommendations from the EUSOMA working group.Eur J Cancer, 46(8):1296–1316, 2010. doi:10.1016/j.ejca.2010.02.015

    work page doi:10.1016/j.ejca.2010.02.015 2010

  4. [4]

    Carney, Diana L

    Patricia A. Carney, Diana L. Miglioretti, Bonnie C. Yankaskas, and et al. Individual and combined effects of age, breast density, and hormone replacement therapy use on the accuracy of screening mammography. Ann Intern Med, 138(3):168–175, 2003. doi:10.7326/0003-4819-138-3-200302040-00008

    work page doi:10.7326/0003-4819-138-3-200302040-00008 2003

  5. [5]

    Fuchsjäger, Paola Clauser, and Ritse M

    Magda Marcon, Michael H. Fuchsjäger, Paola Clauser, and Ritse M. Mann. Esr essentials: screen- ing for breast cancer—general recommendations by EUSOBI.Eur Radiol, 34:6348–6357, 2024. doi:10.1007/s00330-024-10740-5

    work page doi:10.1007/s00330-024-10740-5 2024

  6. [6]

    Trimboli, Alexandra Athanasiou, and et al

    Andrew Evans, Rubina M. Trimboli, Alexandra Athanasiou, and et al. Breast ultrasound: recommen- dations for information to women and referring physicians by the european society of breast imaging. Insights Imaging, 9:449–461, 2018. doi:10.1007/s13244-018-0636-z

    work page doi:10.1007/s13244-018-0636-z 2018

  7. [7]

    Yuchen Zhang, Minghui Xian, H. D. Cheng, Bryar Shareef, Jian Ding, Fei Xu, Kun Huang, Bin Zhang, et al. Busis: A benchmark for breast ultrasound image segmentation.Healthcare, 10(4):729, 2022. doi:10.3390/healthcare10040729

    work page doi:10.3390/healthcare10040729 2022

  8. [8]

    Breast ultrasound tumor classification using a hybrid multitask cnn-transformer network

    Bryar Shareef, Min Xian, Aleksandar Vakanski, and Haotian Wang. Breast ultrasound tumor classification using a hybrid multitask cnn-transformer network. In Hayit Greenspan, Anant Madabhushi, Parvin Mousavi, Septimiu Salcudean, James Duncan, Tanveer Syeda-Mahmood, and Russell Taylor, editors, Medical Image Computing and Computer Assisted Intervention – MIC...

    work page 2023

Show all 57 references
  1. [9]

    D’Orsi, Edward A

    Carol J. D’Orsi, Edward A. Sickles, Ellen B. Mendelson, and Elizabeth A. Morris, editors.ACR BI-RADS Atlas: Breast Imaging Reporting and Data System. American College of Radiology, Reston, VA, 5 edition,

  2. [10]

    Ultrasound radiomics in personalized breast management: Current status and future prospects.Frontiers in Oncology, 12:963612, 2022

    Jionghui Gu and Tian’an Jiang. Ultrasound radiomics in personalized breast management: Current status and future prospects.Frontiers in Oncology, 12:963612, 2022. doi:10.3389/fonc.2022.963612. URL https://pubmed.ncbi.nlm.nih.gov/36059645/

    2022

  3. [11]

    Alison Noble and Dounia Boukerroui

    J. Alison Noble and Dounia Boukerroui. Ultrasound image segmentation: a survey.IEEE Transactions on Medical Imaging, 25(8):987–1010, 2006. doi:10.1109/TMI.2006.877092

    2006

  4. [12]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InMICCAI, volume 9351 ofLNCS, pages 234–241. Springer, 2015. doi:10.1007/978- 3-319-24574-4_28

    2015 doi

  5. [13]

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

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. InDLMIA/ML-CDS@MICCAI, volume 11045 ofLNCS, pages 3–11. Springer, 2018. doi:10.1007/978-3-030-00889-5_1

    2018 doi

  6. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR, pages 770–778, 2016. doi:10.1109/CVPR.2016.90. 12

    2016 doi

  7. [15]

    Weinberger

    Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. InCVPR, pages 4700–4708, 2017. doi:10.1109/CVPR.2017.243

    2017 doi

  8. [16]

    Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert

    Ozan Oktay, Jo Schlemper, Loïc Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y. Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas.arXiv preprint, arXiv:1804.03999, 2018

    2018 arXiv

  9. [17]

    Attention mechanisms in medical image segmentation: A survey.arXiv preprint arXiv:2305.17937, 2023

    Yutong Xie, Bing Yang, Qingbiao Guan, Jianpeng Zhang, Qi Wu, and Yong Xia. Attention mechanisms in medical image segmentation: A survey.arXiv preprint arXiv:2305.17937, 2023

    2023 arXiv

  10. [18]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint, arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2010 arXiv

  11. [19]

    Yuille, and Yuyin Zhou

    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

    2021 arXiv

  12. [20]

    Multi-task breast ultrasound image classifi- cation and segmentation using swin transformer and vmamba models

    Julio Rodriguez, Kuan Huang, and Meng Xu. Multi-task breast ultrasound image classifi- cation and segmentation using swin transformer and vmamba models. In2024 7th Interna- tional Conference on Pattern Recognition and Artificial Intelligence (PRAI), pages 858–863, 2024. doi:10...

    2024

  13. [21]

    Breast ultrasound tumor classification using a hybrid multitask cnn-transformer network

    Bryar Shareef, Min Xian, Aleksandar Vakanski, and Haotian Wang. Breast ultrasound tumor classification using a hybrid multitask cnn-transformer network. InMedical Image Computing and Computer Assisted Intervention – MICCAI 2023, pages 344–353. Springer Nature Switzerland, 2023...

    2023

  14. [22]

    Domain adaptation for medical image analysis: A survey.IEEE Transactions on Biomedical Engineering, 69(3):1173–1185, 2022

    Hao Guan and Mingxia Liu. Domain adaptation for medical image analysis: A survey.IEEE Transactions on Biomedical Engineering, 69(3):1173–1185, 2022. doi:10.1109/TBME.2021.3117407. URL https: //pmc.ncbi.nlm.nih.gov/articles/PMC9011180/

    2022

  15. [23]

    Stan: Small tumor-aware network for breast ultrasound image segmentation

    Bryar Shareef, Min Xian, and Aleksandar Vakanski. Stan: Small tumor-aware network for breast ultrasound image segmentation. In2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), pages 1–5, 2020. doi:10.1109/ISBI45749.2020.9098404. arXiv:2002.01034

    2020

  16. [24]

    Freer, and Min Xian

    Bryar Shareef, Aleksandar Vakanski, Phoebe E. Freer, and Min Xian. Estan: Enhanced small tumor-aware network for breast ultrasound image segmentation.Healthcare, 10(11):2262, 2022. doi:10.3390/healthcare10112262

    2022 doi

  17. [25]

    Incorporating the breast imaging reporting and data system (bi-rads) into deep learning for breast ultrasound.NPJ Breast Cancer, 7:8, 2021

    Yu-Huei Hsieh, Yu-Ting Hung, Yung-Fu Huang, Hsiu-Wei Tsai, Kuan-Jen Chang, Chih-Yu Chen, Chia- Hao Huang, and Po-Hsiang Tsui. Incorporating the breast imaging reporting and data system (bi-rads) into deep learning for breast ultrasound.NPJ Breast Cancer, 7:8, 2021. doi:10.1038...

    2021 doi

  18. [26]

    Incorporating the breast imaging reporting and data system lexicon with a fully convolutional network for malignancy detection on breast ultrasound.Diagnostics, 12(1):66, 2022

    Yung-Hsien Hsieh, Fang-Rong Hsu, Seng-Tong Dai, Hsin-Ya Huang, Dar-Ren Chen, and Wei-Chung Shia. Incorporating the breast imaging reporting and data system lexicon with a fully convolutional network for malignancy detection on breast ultrasound.Diagnostics, 12(1):66, 2022. doi...

    2022 doi

  19. [27]

    Anatomically constrained neural networks (acnn): Application to cardiac image enhance- ment and segmentation.IEEE Transactions on Medical Imaging, 2017

    Ozan Oktay, Enzo Ferrante, Konstantinos Kamnitsas, Mattias Heinrich, Wenjia Bai, Jose Caballero, Stuart Cook, Antonio de Marvao, Timothy Dawes, Declan O’Regan, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Anatomically constrained neural networks (acnn): Application to car...

    2017

  20. [28]

    Boundary loss for highly unbalanced segmentation.Medical Image Analysis, 67:101851, 2021

    Hoel Kervadec, Jalal Bouchtiba, Christian Desrosiers, Eric Granger, Jose Dolz, and Ismail Ben Ayed. Boundary loss for highly unbalanced segmentation.Medical Image Analysis, 67:101851, 2021. doi:10.1016/j.media.2020.101851. 13

    2021

  21. [29]

    Salcudean

    Davood Karimi and Septimiu E. Salcudean. Reducing the hausdorff distance in medical image segmentation with convolutional neural networks.arXiv preprint arXiv:1904.10030, 2019. URL https://arxiv.org/abs/1904.10030

    1904 arXiv

  22. [30]

    Topology-preserving deep image segmentation

    Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-preserving deep image segmentation. InNeurIPS, 2019. URL https://papers.neurips.cc/paper/2019/file/ 2d95666e2649fcfc6e3af75e09f5adb9-Paper.pdf

    2019

  23. [31]

    High-level prior-based loss functions for medical image segmentation: A survey.Computerized Medical Imaging and Graphics, 88: 101852, 2021

    Rayan El Jurdi, Jose Dolz, Christian Desrosiers, and Ismail Ben Ayed. High-level prior-based loss functions for medical image segmentation: A survey.Computerized Medical Imaging and Graphics, 88: 101852, 2021. doi:10.1016/j.compmedimag.2021.101852

    2021

  24. [32]

    Yongjian Yu and Scott T. Acton. Speckle reducing anisotropic diffusion.IEEE Transactions on Image Processing, 11(11):1260–1270, 2002. doi:10.1109/TIP.2002.804276

    2002

  25. [33]

    A critical review and uniformized representation of statistical distributions modeling the envelope of ultrasonic echoes.Ultrasound in Medicine & Biology, 36(7): 1037–1051, 2010

    François Destrempes and Guy Cloutier. A critical review and uniformized representation of statistical distributions modeling the envelope of ultrasonic echoes.Ultrasound in Medicine & Biology, 36(7): 1037–1051, 2010. doi:10.1016/j.ultrasmedbio.2010.04.003

    2010 doi

  26. [34]

    Texture quantified from ultrasound nakagami parametric images is diagnostically relevant for breast lesions.Scientific Reports, 13:10019,

    Siti Muhtadi, Ooi Chuan Leow, Fatimah Abdullah, and et al. Texture quantified from ultrasound nakagami parametric images is diagnostically relevant for breast lesions.Scientific Reports, 13:10019,

  27. [35]

    Christensen, Armando Manduca, Seth Ehrenberg, and et al

    Andrew M. Christensen, Armando Manduca, Seth Ehrenberg, and et al. A systematized review of quantitative ultrasound based acoustic biomarkers and tissue characterization.IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 2024. URL https://pmc.ncbi.nlm.nih...

    2024

  28. [36]

    Segment anything.arXiv preprint arXiv:2304.02643, 2023

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, and et al. Segment anything.arXiv preprint arXiv:2304.02643, 2023

    2023 arXiv

  29. [37]

    URL https://pmc.ncbi.nlm.nih.gov/articles/PMC10285086/

    doi:10.1038/s41598-023-37155-5. URL https://pmc.ncbi.nlm.nih.gov/articles/PMC10285086/

    doi

  30. [38]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection.arXiv preprint arXiv:2303.05499, 2023

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, and et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection.arXiv preprint arXiv:2303.05499, 2023

    2023 arXiv

  31. [39]

    Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159, 2024

    Tianhe Ren, Shilong Liu, Feng Li, and et al. Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159, 2024

    2024 arXiv

  32. [40]

    Timo Lüddecke and Alexander S. Ecker. Image segmentation using text and image prompts.arXiv preprint arXiv:2112.10003, 2021

    2021 arXiv

  33. [41]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, and et al. Microsoft coco: Common objects in context. In ECCV, pages 740–755. Springer, 2014

    2014

  34. [42]

    Berg, and Tamara L

    Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions. InECCV, pages 69–85. Springer, 2016

    2016

  35. [43]

    Junnan Li, Dongxu Li, Steven C. H. Hoi, and Caiming Xiong. Blip: Bootstrapping language–image pre-training for unified vision–language understanding and generation. InProceedings of ICML, volume 162 ofPMLR, pages 12888–12900, 2022

    2022

  36. [44]

    Ultrasound sam adapter: Adapting sam for breast lesion segmentation in ultrasound images.arXiv preprint arXiv:2404.14837, 2024

    Zhengzheng Tu, Le Gu, Xixi Wang, and Bo Jiang. Ultrasound sam adapter: Adapting sam for breast lesion segmentation in ultrasound images.arXiv preprint arXiv:2404.14837, 2024

    2024 arXiv

  37. [45]

    Segment anything in medical images.Nature Communications, 15(1):1022, 2024

    Jun Ma, Yuting He, Feifei Li, and et al. Segment anything in medical images.Nature Communications, 15(1):1022, 2024. doi:10.1038/s41467-024-44824-z. 14

    2024 doi

  38. [46]

    Segment anything model for medical images?arXiv preprint arXiv:2304.14660, 2024

    Yuhao Huang, Xin Yang, Lian Liu, and et al. Segment anything model for medical images?arXiv preprint arXiv:2304.14660, 2024

    2024 arXiv

  39. [47]

    Clip-tnseg: A multi-modal hybrid framework for thyroid nodule segmentation in ultrasound images.arXiv preprint arXiv:2412.05530, 2024

    Xinjie Sun, Boxiong Wei, Yalong Jiang, Liquan Mao, and Qi Zhao. Clip-tnseg: A multi-modal hybrid framework for thyroid nodule segmentation in ultrasound images.arXiv preprint arXiv:2412.05530, 2024

    2024 arXiv

  40. [48]

    Causalclipseg: Unlocking clip’s potential in referring medical image segmentation with causal intervention.arXiv preprint arXiv:2503.15949, 2025

    Yaxiong Chen, Minghong Wei, Zixuan Zheng, and et al. Causalclipseg: Unlocking clip’s potential in referring medical image segmentation with causal intervention.arXiv preprint arXiv:2503.15949, 2025

    2025 arXiv

  41. [49]

    Customized segment anything model for medical image segmentation

    Kaidong Zhang and Dong Liu. Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785, 2023

    2023 arXiv

  42. [50]

    Sarthak Jain and Byron C. Wallace. Attention is not explanation. InProceedings of NAACL-HLT, pages 3543–3556, 2019. URL https://aclanthology.org/N19-1357/

    2019

  43. [51]

    Sanity checks for saliency maps

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. InAdvances in Neural Information Processing Systems (NeurIPS), 2018. URL https://papers.nips.cc/paper/8160-sanity-checks-for-saliency-maps

    2018

  44. [52]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localiza- tion. In2017 IEEE International Conference on Computer Vision (ICCV), pages 618–626, 2017...

    2017 doi

  45. [53]

    U-kan makes strong backbone for medical image segmentation and generation

    Chenxin Li, Xinyu Liu, Wuyang Li, Cheng Wang, Hengyu Liu, Yifan Liu, Zhen Chen, and Yixuan Yuan. U-kan makes strong backbone for medical image segmentation and generation. InProceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conferenc...

    2025 doi

  46. [54]

    Anatosegnet: Anatomy based cnn-transformer network for enhanced breast ultrasound image segmentation

    Meng Xu, Yingfeng Wang, and Kuan Huang. Anatosegnet: Anatomy based cnn-transformer network for enhanced breast ultrasound image segmentation. In2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), pages 1–5. IEEE, 2025. 15

    2025

  47. [55]

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

    Anna Pawłowska, Anna Ćwierz-Pieńkowska, Agnieszka Domalik, Dominika Jaguś, Piotr Kasprzak, Rafał Matkowski, Łukasz Fura, Andrzej Nowicki, and Norbert Żołek. Curated benchmark dataset for ultrasound based breast lesion analysis.Scientific Data, 11(1):148, 2024

    2024

  48. [2013]

    URL https://www.acr.org/Clinical-Resources/Clinical-Tools-and-Reference/ Reporting-and-Data-Systems/BI-RADS

    ISBN 9781559030168. URL https://www.acr.org/Clinical-Resources/Clinical-Tools-and-Reference/ Reporting-and-Data-Systems/BI-RADS

  49. [2023]

    ISBN 978-3-031-43901-8

    Springer Nature Switzerland. ISBN 978-3-031-43901-8

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.