Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

MedGround-R1: Advancing Medical Image Grounding via Spatial-Semantic Rewarded Group Relative Policy Optimization

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

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

Pith's one-line read Group-relative reinforcement learning with box-only supervision achieves state-of-the-art medical image grounding on three public datasets, without chain-of-thought annotations.

desk verdict The MS-CXR SOTA margin likely rides on a MedCLIP reward that may have been pretrained on the same images, so the empirical claim is unproven even though the method idea is sensible. read the letter →

arxiv 2507.02994 v1 pith:5732KLMA submitted 2025-07-01 cs.LG cs.CV

Huihui Xu , Yuanpeng Nie , Hualiang Wang , Ying Chen , Wei Li , Junzhi Ning , Lihao Liu , Hongqiu Wang
show 4 more authors
Lei Zhu Jiyao Liu Xiaomeng Li Junjun He
This is my paper
classification cs.LGcs.CV
keywords medicalimagegroundinggrouprelativepolicyoptimizationreinforcementlearningvision-languagemodelspatial-semanticrewardchain-of-boxreferringexpressioncomprehension
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

Medical image grounding—locating the region a text phrase refers to in a radiology image—normally requires expensive step-by-step reasoning annotations. This paper claims that a reinforcement learning recipe, adapted from large language models, can train a vision-language model for this task using only ground-truth bounding boxes. The model samples several candidate answers per image, scores each one for format compliance, spatial overlap with the true box, and semantic agreement between the cropped region and the phrase, then updates toward the better-scoring completions relative to the group. On three benchmarks (chest X-ray and CT-derived data) the method reports state-of-the-art accuracy and mean IoU, beating the previous best system by 4.48 to 12.06 metric points. The contribution is a training recipe, not a new architecture: the same pretrained vision-language backbone, fine-tuned this way, outperforms its supervised sibling.

What carries the argument

Group Relative Policy Optimization (GRPO): a reinforcement learning algorithm that, for each query, samples a group of G completions from a snapshot of the policy, scores each by scalar rewards, standardizes those scores within the group into advantages, and updates the policy with a clipped, KL-regularized objective toward completions whose relative advantage is positive. In this paper the score is a sum of three rewards: a format reward (regular-expression check for `<think>`/`<answer>` tags and a `[x1, y1, x2, y2]` box), a spatial reward (binary 1 if IoU with the ground-truth box exceeds 0.5, else 0), and a semantic reward (cosine similarity between the cropped predicted region and the referring expression, from a frozen medical contrastive model). The Chain-of-Box template is the other central mechanism: it instructs the model to output its reasoning inside `<think>` tags and to append bounding-box coordinates each time it references a region, converting the hidden spatial attention of the VLM into explicit tokens that the policy gradient can reinforce.

What would settle it

Compute, on MS-CXR, the rank correlation between the frozen semantic model's cosine similarity for a candidate box and that box's IoU with the ground truth, over the spatially-negative completions the paper targets; if the correlation is near zero or negative, the semantic reward is not teaching spatial correctness. A stronger check: retrain the full pipeline on a split whose evaluation images were excluded from the semantic model's pre-training corpus, and see whether the mIoU advantage over the previous best method (79.02 on MS-CXR) survives.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that chain-of-thought annotations are unnecessary for medical image grounding if the reinforcement learning signal is shaped properly. The authors train a vision-language model with Group Relative Policy Optimization (GRPO), which estimates how much better or worse each completion is than its group-mates rather than needing a separate critic. Three rewards define 'better': a format reward for emitting a valid bounding box in the required tags, a spatial reward that gives 1 when the predicted box overlaps the ground truth by more than 0.5 IoU and 0 otherwise, and a semantic reward computed by a frozen medical contrastive vision-language model that scores the cosine similarity between the cropped region and the referring text. The paper further proposes a Chain-of-Box template that asks the model to write its reasoning in `<think>...</think>` tags and, whenever it names a region of interest, to append its bounding-box coordinates, so the visible reasoning trace contains spatial anchors. With these components, the model reports mIoU 79.02 and accuracy 83.12 on MS-CXR, 53.12 and 62.18 on ChestX-ray8, and 60.10 and 74.66 on M3D-RefSeg, surpassing the previous best method on every metric.

Load-bearing premise

The semantic reward assumes that a frozen medical contrastive model's similarity score between a candidate region crop and the referring text is a reliable proxy for whether that region is the right answer; if that similarity is uninformative in the target domain (or was pre-trained on the very images being evaluated), the reward injects noise or bias and the reported gains may not transfer.

Editorial extensions

If this is right

  • Grounding training no longer needs chain-of-thought reasoning annotations; only the ground-truth reference box is required, which reduces annotation cost substantially for new medical grounding tasks.
  • The ablation results show that each reward component carries distinct information: removing the spatial reward collapses mIoU on MS-CXR to 2.66, removing the format reward drops it to 61.22, and removing the semantic reward drops it to 72.14, so the spatial signal is the load-bearing reward and the semantic reward adds a smaller but consistent gain.
  • The same GRPO regime is more stable than supervised fine-tuning on the small MS-CXR training set: SFT overfits (mIoU falls from 68.65 at 1k steps to 56.26 at 3k steps), while GRPO keeps improving to 77.58 at 5k steps and 79.02 with the Chain-of-Box template.
  • Performance improves with larger group sizes (mIoU rises from 77.40 with G=2 to 80.77 with G=8), so the method has headroom if computational cost permits more samples per step.
  • Because the same recipe works across chest X-ray and CT-derived data, the reward design is not tied to one anatomy or imaging modality; the method transfers across the three evaluated benchmarks.

Reading between the lines

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

  • If the gains are real, the same recipe could apply to other dense prediction tasks in medicine where box-level labels exist but fine-grained masks or reasoning annotations do not—for example lesion detection, landmark localization, or region description.
  • An independent check worth running first: the frozen semantic scorer was pre-trained on medical image-text data that overlaps with the test distribution (MS-CXR derives from MIMIC-CXR, a common pre-training corpus), so one should test whether the mIoU advantage over the previous best persists when the scorer has never seen the evaluation images.
  • The paper's spatial reward is binary at IoU 0.5; replacing it with a continuous IoU reward (or adding a secondary fine-grained threshold) would test whether the gap to the upper bound is a reward-shaping artifact, since the binary design gives no gradient incentive to refine a box once it crosses 0.5 IoU.
  • The observed scaling with group size suggests that larger groups could push the benchmarks higher, but the main results use G=4; studying the compute-versus-accuracy frontier at larger G would show where the method plateaus.
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 MedGround-R1, a GRPO-based reinforcement learning method for medical image grounding. It trains Qwen2.5-VL using three rewards: a format reward, a spatial IoU reward against ground-truth boxes, and a semantic reward computed with frozen MedCLIP on the cropped predicted region versus the referring text. It also introduces a Chain-of-Box template that places intermediate bounding boxes inside the <think> reasoning process. Experiments on MS-CXR, ChestX-ray8, and M3D-RefSeg report state-of-the-art mIoU/Acc over seven baselines, with ablations on group size, reward components, and training regime.

Significance. If the empirical claims hold, the paper makes a useful contribution: it shows that GRPO-style RL can replace expensive CoT supervision in medical grounding, using only box annotations and two external reward functions. The method is simple, the ablations are informative, and the use of external rewards (IoU against ground truth and a frozen MedCLIP) keeps the training signal outside the fitted model. Strengths include retraining the BiRD baseline on the same base model and promising code and checkpoint release. The main uncertainty is empirical: the SOTA margin on MS-CXR depends on a MedCLIP semantic reward pretrained on the same source dataset as the evaluation set, and all results are single-run point estimates without error bars.

major comments (4)
  1. [Section 2.3, Table 3, Section 3.1] The semantic reward is computed with a frozen MedCLIP, and MedCLIP is pretrained on MIMIC-CXR, from which MS-CXR is derived. Table 3 shows that removing R_semantic drops MS-CXR mIoU from 79.02 to 72.14, below the BiRD baseline of 73.33; hence the entire reported SOTA margin on MS-CXR rests on this reward. The paper does not report whether MS-CXR evaluation images (or their associated reports) appear in MedCLIP's pretraining, nor does it measure whether MedCLIP similarity to a cropped region correlates with IoU against the ground-truth box. Please provide a leakage check (e.g., pretraining-set overlap analysis) and a spatial-selectivity experiment (e.g., MedCLIP similarity versus IoU over sampled positive and negative boxes), or replace the semantic reward with one that cannot have seen the evaluation images. Without this, the +5.69 mIoU and +7.07 Acc improvements over BiRD on MS-CXR are not established.
  2. [Tables 1-4, Section 3.1] All reported numbers are single-run point estimates without error bars or significance tests. RL training is stochastic, and with G=4 and only 5K training steps, run-to-run variance is unknown. The claims of consistent gains across three datasets, and the G-sweep trend in Table 2, need at least three seeds with mean and standard deviation and, where possible, a test of whether the margin over BiRD is significant. If multiple seeds cannot be run, the empirical claims should be softened accordingly.
  3. [Section 3, Table 1] The comparison protocol is under-specified. It is stated that MS-CXR and ChestX-ray8 are preprocessed following MedRPG, but the exact train/validation/test split sizes are not given, and it is not stated whether all baselines are evaluated on the same split. BiRD is explicitly retrained on Qwen2.5-VL 'same as our setting,' which suggests that other baselines may use numbers from their original papers. Please report the split sizes and either retrain all baselines under the identical protocol or clearly state which numbers are taken from prior work; otherwise Table 1 may conflate model quality with split and training differences.
  4. [Section 2.3, Eq. (1)] The reward combination rule is not specified precisely. The text says 'The sum of these three rewards are then normalized to compute each of their relative advantage A_i,' but Eq. (1) defines a single reward r_i and the normalization across the group. R_format and R_spatial are binary {0,1}, while R_semantic is a cosine similarity; summing them directly assigns different scales and can make the semantic reward dominate or be negligible depending on its range. Please state the exact additive formula, the range of R_semantic (including any scaling), and whether each component is normalized within the group before summation. This is needed to interpret the ablation in Table 3 and to reproduce the method.
minor comments (5)
  1. [Table 4] The header 'mIoU↑IoU↑' appears to be a typo: the second column contains Accuracy values (74.18, 67.50, 73.86, 75.61, 81.09, 83.12) matching the Acc metric used elsewhere, so it should read 'mIoU↑ Acc↑'.
  2. [Section 2.2] The phrase 'Both rewards are implemented by regular expressions' is confusing because only one format reward is described; it should say that both format checks (the think/answer tags and the bounding-box pattern) are implemented via regex.
  3. [Section 3, Datasets] For M3D-RefSeg, please specify how the frontal slice is selected from each 3D volume and whether all 2,778 mask-text-volume triples are used or filtered; this affects reproducibility.
  4. [Section 2.4] The Chain-of-Box template is described as integrating 'referring bounding boxes' into the think process, but these are model-generated boxes rather than ground-truth boxes; please rephrase to avoid ambiguity.
  5. [Abstract and Introduction] The claim of being 'the first RL-based framework for Medical Image Grounding' is strong and unqualified; consider softening it to 'the first, to our knowledge, GRPO-based VLM training without CoT annotations for MIG' unless a systematic literature search is provided.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: all rewards are external functions of ground-truth boxes or a frozen MedCLIP model, and evaluation is on held-out boxes; possible MedCLIP pretraining overlap is a data-leakage concern, not definitional circularity.

full rationale

The paper's derivation chain is a standard RL fine-tuning procedure: Qwen2.5-VL is trained with GRPO using three rewards, R_format (regular-expression format check), R_spatial (IoU against the ground-truth box, thresholded at 0.5), and R_semantic (cosine similarity between a frozen MedCLIP crop feature and the referring-expression feature). The evaluation metrics are mIoU and Acc (IoU > 0.5) computed on test examples. Although R_spatial uses the same IoU>0.5 criterion as the Acc metric, this is direct optimization of the target metric on training annotations, not a fitted parameter being renamed as a prediction; the training and evaluation quantities are not the same data points, and no equation in the paper defines a reported result in terms of a fitted input. The semantic reward is computed with a frozen external model (MedCLIP), not learned from the evaluation data, and the paper is transparent that this reward is ablatable: without it, MS-CXR mIoU drops from 79.02 to 72.14. Citations to related work, including OphCLIP and MedCLIP, are contextual rather than load-bearing uniqueness claims. The concern that MedCLIP may have been pretrained on MIMIC-CXR, from which MS-CXR is derived, would be a data-leakage or generalization-risk issue, but it is not a circularity of the kind where a prediction is equivalent to its input by construction. Accordingly, the paper merits a circularity score of 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard GRPO machinery, a handful of hyperparameters, and the domain-specific assumption that MedCLIP similarity is a useful training signal. No new physical entities are introduced. The equal-weight reward sum is an untuned design choice.

free parameters (5)
  • IoU threshold for spatial reward = 0.5
    Threshold above which the spatial reward is 1; matches the evaluation Acc threshold (IoU>0.5). Not tuned independently.
  • Group size G = 4 (main); 2, 6, 8 in ablation
    Number of sampled completions per step in GRPO; affects performance and compute.
  • KL coefficient beta = 0.04
    Weight for KL divergence against the reference policy.
  • Reward combination weights = 1,1,1 (implicit)
    The three rewards are simply summed with no stated weights or tuning.
  • Training schedule = 5K steps, lr 1e-6, batch 1 x grad accum 2
    Hyperparameters chosen for training; only G is ablated.
assumptions (4)
  • standard math GRPO objective (Eq. 4) is a valid policy gradient estimator with clipping and KL penalty.
    Borrowed from DeepSeek-R1 and DeepSeek-Math.
  • domain assumption MedCLIP cosine similarity between a cropped ROI and the referring expression is a meaningful proxy for semantic consistency in the medical domain.
    The semantic reward's usefulness depends on this; no validation of MedCLIP's alignment on the target datasets is provided.
  • domain assumption Ground truth bounding boxes in MS-CXR, ChestX-ray8, and M3D-RefSeg are accurate and sufficient for training and evaluation.
    All boxes are used as both training targets and evaluation labels; the ChestX-ray8 boxes are category-level labels.
  • domain assumption The 2D frontal-slice conversion of M3D-RefSeg preserves the localization signal for grounding.
    The paper extracts a slice from each 3D volume without specifying which axis; this conversion is used for both training and evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MedGround-R1: Advancing Medical Image Grounding via Spatial-Semantic Rewarded Group Relative Policy Optimization." pith.science (2026). https://pith.science/paper/5732KLMA

@misc{pith2026250702994,
  author       = {Pith},
  title        = {Pith review of: MedGround-R1: Advancing Medical Image Grounding via Spatial-Semantic Rewarded Group Relative Policy Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5732KLMA}},
  note         = {Machine review of arXiv:2507.02994}
}
read the original abstract

Medical Image Grounding (MIG), which involves localizing specific regions in medical images based on textual descriptions, requires models to not only perceive regions but also deduce spatial relationships of these regions. Existing Vision-Language Models (VLMs) for MIG often rely on Supervised Fine-Tuning (SFT) with large amounts of Chain-of-Thought (CoT) reasoning annotations, which are expensive and time-consuming to acquire. Recently, DeepSeek-R1 demonstrated that Large Language Models (LLMs) can acquire reasoning abilities through Group Relative Policy Optimization (GRPO) without requiring CoT annotations. In this paper, we adapt the GRPO reinforcement learning framework to VLMs for Medical Image Grounding. We propose the Spatial-Semantic Rewarded Group Relative Policy Optimization to train the model without CoT reasoning annotations. Specifically, we introduce Spatial-Semantic Rewards, which combine spatial accuracy reward and semantic consistency reward to provide nuanced feedback for both spatially positive and negative completions. Additionally, we propose to use the Chain-of-Box template, which integrates visual information of referring bounding boxes into the <think> reasoning process, enabling the model to explicitly reason about spatial regions during intermediate steps. Experiments on three datasets MS-CXR, ChestX-ray8, and M3D-RefSeg demonstrate that our method achieves state-of-the-art performance in Medical Image Grounding. Ablation studies further validate the effectiveness of each component in our approach. Code, checkpoints, and datasets are available at https://github.com/bio-mlhui/MedGround-R1

Figures

Figures reproduced from arXiv: 2507.02994 by the authors.

Figure 1
Figure 1. A Training Step of the Proposed Spatial-Semantic Group Relative [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparisons on MS-CXR. The blue dashed boxes are ground truth bounding boxes. “abnormal opacity in the basilar right hemithorax is likely atelectasis involving the right lower and middle lobes” “Consolidation in the mid left lung continues to expand, worsening pneumonia” “patc h y g r ou n d￾glass opacity in the right lung base” “The apical component has slightly decreased” Ours BiRD ChEX CausalCLIPSeg M… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. EventRR: Event Referential Reasoning for Referring Video Object Segmentation

    cs.CV 2025-08 conditional novelty 7.0 of 10

    EventRR builds a Referential Event Graph from AMR parsing of the referring expression and uses graph-guided temporal reasoning over detector queries to select and segment the referent, reporting state-of-the-art resul...

  2. From Failure to Feedback: Group Revision Unlocks Hard Cases in Object-Level Grounding

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    A group-revision paradigm for GRPO-based RL fine-tuning of VLMs converts failure responses into improvement signals that refine rewards and advantages, yielding gains on referring segmentation, REC, and counting benchmarks.

Reference graph

Works this paper leans on

30 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    arXiv preprint arXiv:2404.00578 (2024)

    Bai, F., Du, Y., Huang, T., Meng, M.Q.H., Zhao, B.: M3d: Advancing 3d medical image analysis with multi-modal large language models. arXiv preprint arXiv:2404.00578 (2024)

    arXiv 2024

  2. [2]

    5-vl technical report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al.: Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    arXiv 2025

  3. [3]

    In: European conference on computer vision

    Boecking, B., Usuyama, N., Bannur, S., Castro, D.C., Schwaighofer, A., Hyland, S., Wetscherek, M., Naumann, T., Nori, A., Alvarez-Valle, J., et al.: Making the most of text semantics to improve biomedical vision–language processing. In: European conference on computer vision. pp. 1–21. Springer (2022)

    2022

  4. [4]

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

    Chen, Y., Wei, M., Zheng, Z., Hu, J., Shi, Y., Xiong, S., Zhu, X.X., Mou, L.: Causalclipseg: Unlocking clip’s potential in referring medical image segmentation with causal intervention. In: International Conference on Medical Image Comput- ing and Computer-Assisted Intervention. pp. 77–87. Springer (2024)

    work page 2024

  5. [5]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Chen, Y., Wang, G., Ji, Y., Li, Y., Ye, J., Li, T., Hu, M., Yu, R., Qiao, Y., He, J.: Slidechat: A large vision-language assistant for whole-slide pathology image understanding. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5134–5143 (2025)

    work page 2025

  6. [6]

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

    Chen, Z., Zhou, Y., Tran, A., Zhao, J., Wan, L., Ooi, G.S.K., Cheng, L.T.E., Thng, C.H., Xu, X., Liu, Y., et al.: Medical phrase grounding with region-phrase context contrastive alignment. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 371–381. Springer (2023)

    2023

  7. [7]

    arXiv preprint arXiv:2501.12948 (2025)

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al.: Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025)

    arXiv 2025

  8. [8]

    OphCLIP: Hierarchical Retrieval-Augmented Learning for Ophthalmic Surgical Video-Language Pretraining

    Hu, M., Yuan, K., Shen, Y., Tang, F., Xu, X., Zhou, L., Li, W., Chen, Y., Xu, Z., Peng, Z., et al.: Ophclip: Hierarchical retrieval-augmented learning for ophthalmic surgical video-language pretraining. arXiv preprint arXiv:2411.15421 (2024)

    work page Pith review arXiv 2024

Show all 30 references
  1. [9]

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

    Huang, X., Huang, H., Shen, L., Yang, Y., Shang, F., Liu, J., Liu, J.: A refer- and-ground multimodal large language model for biomedicine. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 399–409. Springer (2024)

    2024

  2. [10]

    IEEE Transactions on Medical Imaging (2024)

    Huang, X., Li, H., Cao, M., Chen, L., You, C., An, D.: Cross-modal conditioned re- construction for language-guided medical image segmentation. IEEE Transactions on Medical Imaging (2024)

    2024

  3. [11]

    Scientific data6(1), 317 (2019)

    Johnson, A.E., Pollard, T.J., Berkowitz, S.J., Greenbaum, N.R., Lungren, M.P., Deng, C.y., Mark, R.G., Horng, S.: Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data6(1), 317 (2019)

    2019

  4. [12]

    Pattern Recognition p

    Li, Y., Wang, H., Duan, Y., Zhang, J., Li, X.: A closer look at the explainability of contrastive language-image pre-training. Pattern Recognition p. 111409 (2025)

    2025

  5. [13]

    IEEE transactions on medical imaging43(1), 96–107 (2023)

    Li, Z., Li, Y., Li, Q., Wang, P., Guo, D., Lu, L., Jin, D., Zhang, Y., Hong, Q.: Lvit: language meets vision transformer in medical image segmentation. IEEE transactions on medical imaging43(1), 96–107 (2023)

    2023

  6. [14]

    In: European Conference on Computer Vision

    Müller, P., Kaissis, G., Rueckert, D.: Chex: Interactive localization and region description in chest x-rays. In: European Conference on Computer Vision. pp. 92–

  7. [15]

    arXiv preprint arXiv:1707.06347 (2017)

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

  8. [16]

    arXiv preprint arXiv:2402.03300 (2024)

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024)

    2024 arXiv

  9. [17]

    IEEE Transactions on Big Data9(4), 1102–1115 (2023)

    Tian, X., Ng, W.W., Xu, H.: Deep incremental hashing for semantic image retrieval with concept drift. IEEE Transactions on Big Data9(4), 1102–1115 (2023)

    2023

  10. [18]

    arXiv preprint arXiv:2409.04356 (2024)

    Wang, H., Chen, Y., Chen, W., Xu, H., Zhao, H., Sheng, B., Fu, H., Yang, G., Zhu, L.: Serp-mamba: Advancing high-resolution retinal vessel segmentation with selective state-space model. arXiv preprint arXiv:2409.04356 (2024)

    2024 arXiv

  11. [19]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Wang, H., Wang, W., Zhou, H., Xu, H., Wu, S., Zhu, L.: Language-driven interac- tive shadow detection. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 5527–5536 (2024)

    2024

  12. [20]

    arXiv preprint arXiv:2410.18387 (2024)

    Wang, L., Wang, H., Yang, H., Mao, J., Yang, Z., Shen, J., Li, X.: Interpretable bilingual multimodal large language model for diverse biomedical tasks. arXiv preprint arXiv:2410.18387 (2024)

    2024 arXiv

  13. [21]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., Summers, R.M.: Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classi- fication and localization of common thorax diseases. In: Proceedings of the IEEE conference on computer vision and patte...

    2017

  14. [22]

    In: Proceedings of the Conference on Empirical Methods in Natural Language Processing

    Wang, Z., Wu, Z., Agarwal, D., Sun, J.: Medclip: Contrastive learning from un- paired medical images and text. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing. vol. 2022, p. 3...

    2022

  15. [23]

    Radiology: Artificial Intelligence 5(5), e230024 (2023)

    Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: robust segmen- tation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5), e230024 (2023)

    2023

  16. [24]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Wu, H., Yang, Y., Xu, H., Wang, W., Zhou, J., Zhu, L.: Rainmamba: Enhanced locality learning with state space models for video deraining. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 7881–7890 (2024)

    2024

  17. [25]

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

    Xu, H., Yang, Y., Aviles-Rivero, A.I., Yang, G., Qin, J., Zhu, L.: Lgrnet: Local- global reciprocal network for uterine fibroid segmentation in ultrasound videos. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 667–677. Springer (2024)

    2024

  18. [26]

    Xu, H., Zhu, L.: Online agglomerative pooling for scalable self-supervised universal segmentation (2024), https://openreview.net/forum?id=d32d9fE5lG

    2024

  19. [27]

    Physics in Medicine & Biology66(17), 175003 (2021)

    Yu, L., Zhang, Z., Li, X., Ren, H., Zhao, W., Xing, L.: Metal artifact reduction in 2d ct images with self-supervised cross-domain learning. Physics in Medicine & Biology66(17), 175003 (2021)

    2021

  20. [28]

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

    Zhong, Y., Xu, M., Liang, K., Chen, K., Wu, M.: Ariadne’s thread: Using text prompts to improve segmentation of infected areas from chest x-ray images. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 724–733. Springer (2023)

    2023

  21. [29]

    arXiv preprint arXiv:2404.06798 (2024)

    Zou, K., Bai, Y., Chen, Z., Zhou, Y., Chen, Y., Ren, K., Wang, M., Yuan, X., Shen, X., Fu, H.: Medrg: Medical report grounding with multi-modal large language model. arXiv preprint arXiv:2404.06798 (2024)

    2024 arXiv

  22. [111]

    Springer (2024) MedGround-R1 11

    2024

Pith tools

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