pith. sign in

arxiv: 2606.24570 · v2 · pith:GDYO7NRHnew · submitted 2026-06-23 · 💻 cs.CV

Jolia: Concept-Level Vision-Language Alignment for 3D CT Contrastive Learning

Julien Khlaut , Charles Corbi\`ere , Baptiste Callard , Amaury Prat , Leo Butsanets , Antoine Saporta , Th\'eo Danielou , Leo Machado
show 4 more authors
Korentin Le Floch Tom Boeken Pierre Manceron Corentin Dancette
This is my paper

Pith reviewed 2026-06-26 00:27 UTC · model grok-4.3

classification 💻 cs.CV
keywords 3D CTvision-language modelscontrastive learningconcept querieslocalized alignmentmedical foundation modelsradiological reportscross-attention
0
0 comments X

The pith

Adding localized concept alignments to contrastive pretraining improves 3D CT vision-language models over global CLIP baselines.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Vision-language contrastive pretraining for 3D medical images typically encodes each scan and report as a single global token, which risks losing details across multiple organs and the structured sections of long radiological reports. The paper augments this global alignment with independent localized alignments, one per concept such as an anatomical region, by splitting reports into concept-specific sections and learning cross-attention queries that pool matching image features. These queries require no segmentation masks or spatial labels. The resulting Jolia model, trained on chest and abdominal CT, outperforms a CLIP baseline on findings classification, report generation, and cross-center transfer while reaching new state-of-the-art results on public benchmarks. A sympathetic reader would care because finer-grained alignments could preserve clinically relevant details that global encodings discard.

Core claim

The paper claims that augmenting CLIP-style global alignment with a set of localized alignments—one per concept—where reports are split into concept-specific sections and cross-attention queries learn to pool matching image features independently for each concept, produces 3D CT foundation models like Jolia that consistently outperform global-only baselines on findings classification, report generation, and cross-center transfer, while incidentally generating attention maps focused on each concept for built-in spatial interpretability.

What carries the argument

ConQuer (Concept Queries): a set of cross-attention queries that perform independent localized alignments between image features and concept-specific sections of the radiological report.

If this is right

  • Jolia outperforms a CLIP baseline on findings classification tasks.
  • The model improves performance on radiological report generation.
  • Cross-center transfer learning results improve over the global baseline.
  • New state-of-the-art results are achieved across multiple public 3D CT benchmarks.
  • Each learned query produces attention maps focused on its corresponding anatomical concept.

Where Pith is reading between the lines

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

  • The same splitting-and-query approach could be tested on other 3D modalities such as MRI without requiring new spatial annotations.
  • Attention maps from the queries might be used directly for clinical verification of model attention on specific organs.
  • Extending the method to finer concepts like lesions or pathologies would require only report sectioning rather than new labels.
  • The pattern of replacing one global alignment with multiple semantic alignments may apply to contrastive learning on other structured multimodal data.

Load-bearing premise

That splitting radiological reports into concept-specific sections and training independent cross-attention queries without any spatial supervision will produce meaningful localized alignments that improve downstream performance.

What would settle it

Training an otherwise identical model without the concept queries and finding no gain or a loss in accuracy on findings classification, report generation metrics, or cross-center transfer would falsify the claimed benefit of the localized alignments.

Figures

Figures reproduced from arXiv: 2606.24570 by Amaury Prat, Antoine Saporta, Baptiste Callard, Charles Corbi\`ere, Corentin Dancette, Julien Khlaut, Korentin Le Floch, Leo Butsanets, Leo Machado, Pierre Manceron, Th\'eo Danielou, Tom Boeken.

Figure 1
Figure 1. Figure 1: Overview of ConQuer. We decompose both the image (via learnable cross-attention queries) and the report (via concept-specific sections produced by an LLM) into a set of concept-level representations, and apply contrastive alignment independently for each concept in addition to the standard global LCLIP(z[CLS], t). A fundamental limitation of standard CLIP-style alignment, however, is that both the image an… view at source ↗
Figure 2
Figure 2. Figure 2: Evaluation protocol for findings classification (linear probing and zero-shot): we concatenate [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Zero-shot evaluation results. (a) Per-template and aggregate zero-shot AUROC on the test set. Each cell reports the macro-AUROC obtained with a single prompt template pair (columns 1–8) or their aggregate (rightmost column), evaluated on CT-RATE (top) and Stanford Abd CT (bottom). (b) Ablation study on the number of prompt templates (r) in long-form zero-shot evaluation. 4.4 Radiology Report Generation We … view at source ↗
Figure 4
Figure 4. Figure 4: Organ-level attention maps. Cross-attention maps for different organ queries show that Jolia models learn to attend to anatomically meaningful regions without explicit supervision. 5 Conclusion We introduced ConQuer (Concept Queries), an image–text contrastive pretraining method that augments global CLIP alignment with localized, concept-level alignments learned from concept￾specific report sections via cr… view at source ↗
Figure 5
Figure 5. Figure 5: Example of the LLM-based report-splitting pipeline. A raw chest CT Findings report (left, paraphrased from a public CT-RATE study) is decomposed into atomic, single-sentence obser￾vations, each tagged with a finding_category (in grey, in brackets) and an organ drawn from our taxonomy. Sentences sharing an anatomical entry are concatenated into one organ section (right) before being embedded by the frozen t… view at source ↗
Figure 6
Figure 6. Figure 6: Aggregation converges to a stable zero-shot estimate. Test macro-AUC against the number of aggregated templates p on CT-RATE (left) and Merlin-Abd-CT (right). Solid lines show the aggregation protocol; dashed and dotted lines show the val-best and val-worst single-template AUCs, with the shaded tube denoting their envelope. Curves stabilise by p ≈ 4–5 on both datasets and no model collapses at p = 1, suppo… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of report generation on two held-out Merlin-Abd-CT abdomi￾nal CT studies: a mostly-normal study (a) and a complex post-pelvic-exenteration study (b). Green marks findings supported by the radiologist ground truth; red marks hallucinations; plain text is filler / normal phrasing. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Principal Component Analysis Visualization. First three non-trivial PCA components of the high-resolution feature maps of the same volume mapped to RGB channels. Jolia models produce the most spatially structured and anatomically coherent representations across all views. Pancreas Spleen Axial Coronal Sagittal [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Failure cases in organ-level attention. Cross-attention maps for the pancreas and spleen demonstrate lower localization precision compared to other organs. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
read the original abstract

Vision-language contrastive pretraining has become the dominant recipe for 3D medical foundation models, leveraging the large volumes of paired scans and reports produced in clinical practice. However, medical images usually span dozens of organs, and radiological reports are much longer than typical natural image captions and are composed of multiple structured sections. CLIP-style pretraining compresses this structure by encoding each modality into a single global token, at the risk of losing important details. We introduce ConQuer (Concept Queries), an image-text pretraining method that augments CLIP's global alignment with a set of localized alignments, one per concept. ConQuer splits the report into concept-specific sections and learns cross-attention queries that pool the matching image features without using any segmentation mask or spatial supervision. Contrastive learning is then applied independently for each concept. Concepts can be any unit of semantic localization; here, they are anatomical regions, one query per organ or gross body region. As a byproduct, each query learns attention maps focused on its concept, providing built-in spatial interpretability. We use ConQuer to train Jolia, a 3D CT foundation model on chest and abdominal CT. Jolia consistently outperforms a CLIP baseline on findings classification, report generation, and cross-center transfer, and sets a new state of the art across multiple public benchmarks. Jolia's weights are available at https://huggingface.co/raidium/Jolia

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces Jolia, a 3D CT vision-language foundation model trained with ConQuer (Concept Queries). ConQuer augments standard global CLIP alignment by splitting radiological reports into concept-specific sections (anatomical regions), learning one cross-attention query per concept to pool matching image features, and applying independent per-concept contrastive losses. No segmentation masks or spatial supervision are used. The resulting model is claimed to outperform a CLIP baseline on findings classification, report generation, and cross-center transfer while setting new state-of-the-art results across multiple public benchmarks.

Significance. If the per-concept queries successfully induce localized alignments that drive the reported gains, the approach would address a key limitation of global encodings when handling multi-organ 3D CT scans and structured multi-section reports, while also providing built-in spatial interpretability via attention maps. This could represent a meaningful advance for medical foundation models by enabling more structured contrastive pretraining without additional annotations.

major comments (2)
  1. [Abstract] Abstract: the central claim that Jolia 'consistently outperforms a CLIP baseline' and 'sets a new state of the art' is unsupported by any quantitative results, ablation details, or experimental setup. This is load-bearing because the soundness of the outperformance (and thus the value of ConQuer) cannot be evaluated from the given information.
  2. [Abstract] Abstract (ConQuer description): the method asserts that independent cross-attention queries will produce meaningful localized alignments for each anatomical concept, yet supplies no mechanism (e.g., orthogonality, attention regularization, reconstruction loss, or spatial term) to prevent collapse to global or shared representations. This directly affects the weakest assumption underlying the claimed gains over CLIP, as performance differences could instead arise from extra parameters or report preprocessing.
minor comments (1)
  1. The Hugging Face weight link is mentioned only in the abstract; ensure it receives a formal citation and is referenced in the main text and reproducibility section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments on the abstract below and have revised the manuscript to incorporate quantitative support and methodological clarifications.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that Jolia 'consistently outperforms a CLIP baseline' and 'sets a new state of the art' is unsupported by any quantitative results, ablation details, or experimental setup. This is load-bearing because the soundness of the outperformance (and thus the value of ConQuer) cannot be evaluated from the given information.

    Authors: We agree that the abstract should provide more concrete support. In the revised manuscript we will add key quantitative results (e.g., AUC gains on multi-label classification, BLEU/ROUGE improvements on report generation, and transfer performance deltas) together with a one-sentence description of the evaluation protocol. This keeps the abstract concise while making the claims evaluable. revision: yes

  2. Referee: [Abstract] Abstract (ConQuer description): the method asserts that independent cross-attention queries will produce meaningful localized alignments for each anatomical concept, yet supplies no mechanism (e.g., orthogonality, attention regularization, reconstruction loss, or spatial term) to prevent collapse to global or shared representations. This directly affects the weakest assumption underlying the claimed gains over CLIP, as performance differences could instead arise from extra parameters or report preprocessing.

    Authors: The abstract is intentionally high-level. The full paper contains attention-map visualizations and controlled ablations showing that the per-concept queries produce distinct, organ-focused alignments rather than collapsing, even without explicit regularization. We will revise the abstract to note this empirical outcome. We maintain that the gains are not explained by parameter count alone, as our ablations match capacity; however, we accept the need to clarify the assumption in the abstract and will do so. revision: partial

Circularity Check

0 steps flagged

No circularity: method defined independently; gains are empirical, not forced by construction.

full rationale

The paper defines ConQuer by splitting reports into concept sections, introducing per-concept cross-attention queries, and applying independent contrastive losses without spatial labels. These choices are explicit design decisions, not derived from or equivalent to the target performance metrics. Downstream improvements on classification, generation, and transfer are reported as experimental outcomes on public benchmarks, not as quantities that reduce to the training inputs by construction. No self-citation chains, fitted parameters renamed as predictions, or ansatzes smuggled via prior work appear in the provided text. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are detailed beyond the high-level method description.

invented entities (1)
  • ConQuer no independent evidence
    purpose: Augment CLIP with per-concept localized alignments
    New method name and procedure introduced in the work

pith-pipeline@v0.9.1-grok · 5837 in / 1035 out tokens · 21932 ms · 2026-06-26T00:27:56.064432+00:00 · methodology

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

34 extracted references · 4 linked inside Pith

  1. [1]

    On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arber, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021

    Pith/arXiv arXiv 2021

  2. [2]

    Learning transferable visual models from natural language supervision.International Conference on Machine Learning (ICML), 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision.International Conference on Machine Learning (ICML), 2021

    2021

  3. [3]

    Is CLIP ideal? No

    Raphi Kang, Yue Song, Georgia Gkioxari, and Pietro Perona. Is CLIP ideal? No. Can we fix it? Yes!arXiv preprint arXiv:2503.08723, 2025

    arXiv 2025

  4. [4]

    Large-scale and fine-grained vision-language pre-training for enhanced ct image understanding.International Conference on Learning Representations (ICLR), 2025

    Zhongyi Shui, Jianpeng Zhang, Weiwei Cao, et al. Large-scale and fine-grained vision-language pre-training for enhanced ct image understanding.International Conference on Learning Representations (ICLR), 2025

    2025

  5. [5]

    Totalfm: An organ-separated framework for 3d-ct vision foundation models.arXiv preprint arXiv:2601.00260, 2026

    Kohei Yamamoto and Tomohiro Kikuchi. Totalfm: An organ-separated framework for 3d-ct vision foundation models.arXiv preprint arXiv:2601.00260, 2026

    arXiv 2026

  6. [6]

    Ct-glip: 3d grounded language-image pretrain- ing with ct scans and radiology reports for full-body scenarios.arXiv preprint arXiv:2404.15272, 2024

    Jingyang Lin, Yingda Xia, Jianpeng Zhang, et al. Ct-glip: 3d grounded language-image pretrain- ing with ct scans and radiology reports for full-body scenarios.arXiv preprint arXiv:2404.15272, 2024

    arXiv 2024

  7. [8]

    Lungren, Curtis P

    Shih-Cheng Huang, Zepeng Huo, Ethan Steinberg, Chia-Chun Chiang, Matthew P. Lungren, Curtis P. Langlotz, Serena Yeung, Nigam H. Shah, and Jason A. Fries. INSPECT: A multimodal dataset for pulmonary embolism diagnosis and prognosis. InAdvances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2023

    2023

  8. [9]

    Merlin: A vision language foundation model for 3d computed tomography.Nature, 2024

    Louis Blankemeier, Joseph Paul Cohen, Ashwin Kumar, Dave Van Veen, Syed Jamal Safdar Gardezi, Magdalini Paez, Curt Dorfman, Vishwanatha Venugopal, et al. Merlin: A vision language foundation model for 3d computed tomography.Nature, 2024

    2024

  9. [10]

    Lungren, Tristan Naumann, Sheng Wang, and Hoifung Poon

    Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, Cliff Wong, Andrea Tupini, Yu Wang, Matt Mazzola, Swadheen Shukla, Lars Liden, Jianfeng Gao, Matthew P. Lungren, Tristan Naumann, Sheng Wang, and Hoifung Poon. BiomedCLIP: a multimodal biomedical foundation model pretrained from f...

    Pith/arXiv arXiv 2024

  10. [11]

    Noel C. F. Codella, Ying Jin, Shrey Jain, Yu Gu, Ho Hin Lee, Asma Ben Abacha, Alberto Santamaria-Pang, Will Guyman, Naiteek Sangani, Sheng Zhang, Hoifung Poon, Stephanie Hyland, Shruthi Bannur, Javier Alvarez-Valle, Xue Li, John Garrett, Alan McMillan, Gaurav Rajguru, Madhu Maddi, Nilesh Vijayrania, Rehaan Bhimai, Nick Mecklenburg, Rupal Jain, Daniel Hols...

    arXiv 2024

  11. [12]

    Steiner, Can Kirmizibayrak, Rory Pilgrim, Daniel Golden, and Lin Yang

    Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, Cían Hughes, Charles Lau, Justin Chen, Fereshteh Mahvar, Liron Yatziv, Tiffany Chen, Bram Sterling, Stefanie Anna Baby, Susanna Maria Baby, Jeremy Lai, Samuel Schmidgall, Lu Yang, Kejia Chen, Per Bjornsson, Shashir Reddy, Ryan Br...

    Pith/arXiv arXiv 2025

  12. [13]

    Ct-clip, a pre-trained text-image model for 3d chest ct image and radiology report analysis.arXiv preprint arXiv:2403.17834, 2024

    Ibrahim Ethem Hamamci, Sezgin Er, and Bjoern Menze. Ct-clip, a pre-trained text-image model for 3d chest ct image and radiology report analysis.arXiv preprint arXiv:2403.17834, 2024

    arXiv 2024

  13. [14]

    Pillar-0: A new frontier for radiology foundation models.arXiv preprint arXiv:2511.17803, 2025

    Akshay S Chaudhari et al. Pillar-0: A new frontier for radiology foundation models.arXiv preprint arXiv:2511.17803, 2025

    arXiv 2025

  14. [15]

    Comprehensive language-image pre-training for 3d medical image understanding.arXiv preprint arXiv:2510.15042, 2025

    Tassilo Wald, Ibrahim Ethem Hamamci, Yuan Gao, Sam Bond-Taylor, Harshita Sharma, Max- imilian Ilse, Cynthia Lo, Olesya Melnichenko, Anton Schwaighofer, Noel CF Codella, et al. Comprehensive language-image pre-training for 3d medical image understanding.arXiv preprint arXiv:2510.15042, 2025

    arXiv 2025

  15. [16]

    Kann, Andriy Fedorov, Raymond H

    Suraj Pai, Ibrahim Hadzic, Dennis Bontempi, Keno Bressem, Benjamin H. Kann, Andriy Fedorov, Raymond H. Mak, and Hugo J. W. L. Aerts. Vision foundation models for computed tomography, 2025

    2025

  16. [17]

    Scaling self-supervised and cross-modal pretraining for volumetric ct transformers

    Cris Claessens, Christiaan Viviers, Giacomo D’Amicantonio, Egor Bondarev, and Fons van der Sommen. Scaling self-supervised and cross-modal pretraining for volumetric ct transformers. arXiv preprint arXiv:2511.17209, 2025

    arXiv 2025

  17. [18]

    Curia: A multi- modal foundation model for radiology.arXiv preprint arXiv:2509.06830, 2025

    Corentin Dancette, Julien Khlaut, Antoine Saporta, Helene Philippe, Elodie Ferreres, Baptiste Callard, Théo Danielou, Léo Alberge, Léo Machado, Daniel Tordjman, et al. Curia: A multi- modal foundation model for radiology.arXiv preprint arXiv:2509.06830, 2025

    arXiv 2025

  18. [19]

    Curia-2: Scaling self-supervised learning for radiology foundation models.arXiv preprint arXiv:2604.01987, 2026

    Antoine Saporta, Baptiste Callard, Corentin Dancette, Julien Khlaut, Charles Corbière, Leo Butsanets, Amaury Prat, and Pierre Manceron. Curia-2: Scaling self-supervised learning for radiology foundation models.arXiv preprint arXiv:2604.01987, 2026

    arXiv 2026

  19. [20]

    FILIP: Fine-grained interactive language-image pre-training

    Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained interactive language-image pre-training. InInternational Conference on Learning Representations (ICLR), 2022

    2022

  20. [21]

    DenseCLIP: Language-guided dense prediction with context-aware prompting

    Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guangyu Huang, Jie Zhou, and Jiwen Lu. DenseCLIP: Language-guided dense prediction with context-aware prompting. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  21. [22]

    Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition

    Shih-Cheng Huang, Liyue Shen, Matthew P Lungren, and Serena Yeung. Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. IEEE/CVF International Conference on Computer Vision (ICCV), 2021

    2021

  22. [23]

    Multi- granularity cross-modal alignment for generalized medical visual representation learning

    Fuying Wang, Yuyin Zhou, Shujun Wang, Varut Vardhanabhuti, and Lequan Yu. Multi- granularity cross-modal alignment for generalized medical visual representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 11

    2022

  23. [24]

    RadSAM: Segmenting 3D radiological images with a 2D promptable model.arXiv preprint arXiv:2504.20837, 2025

    Julien Khlaut, Elodie Ferreres, Daniel Tordjman, Hélène Philippe, Tom Boeken, Pierre Manceron, and Corentin Dancette. RadSAM: Segmenting 3D radiological images with a 2D promptable model.arXiv preprint arXiv:2504.20837, 2025

    arXiv 2025

  24. [25]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean Conference on Computer Vision (ECCV), 2020

    2020

  25. [26]

    Perceiver: General perception with iterative attention.International Conference on Machine Learning (ICML), 2021

    Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and Joao Carreira. Perceiver: General perception with iterative attention.International Conference on Machine Learning (ICML), 2021

    2021

  26. [27]

    BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational Conference on Machine Learning (ICML), 2023

    2023

  27. [28]

    Atlas: Multi-scale attention improves long context image modeling.ArXiv, abs/2503.12355, 2025

    Kumar Krishna Agrawal, Long Lian, Longchao Liu, Natalia Harguindeguy, Boyi Li, Alexan- der G Bick, Maggie Chung, Trevor Darrell, and Adam Yala. Atlas: Multi-scale attention improves long context image modeling.ArXiv, abs/2503.12355, 2025

    arXiv 2025

  28. [29]

    Qwen3 embedding: Advancing text embedding and reranking through foundation models.arXiv preprint arXiv:2506.05176, 2025

    Yanzhao Zhang, Mingxin Li, Dingkun Long, et al. Qwen3 embedding: Advancing text embedding and reranking through foundation models.arXiv preprint arXiv:2506.05176, 2025

    Pith/arXiv arXiv 2025

  29. [30]

    Qwen3.5: Accelerating productivity with native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Accelerating productivity with native multimodal agents, February 2026

    2026

  30. [31]

    GREEN: Generative radiology report evaluation and error notation

    Sophie Ostmeier, Justin Xu, Zhihong Chen, Maya Varma, Louis Blankemeier, Christian Blueth- gen, Arne Edward Michalson, Michael Moseley, Curtis Langlotz, Akshay S Chaudhari, and Jean-Benoit Delbrouck. GREEN: Generative radiology report evaluation and error notation. In Findings of the Association for Computational Linguistics: EMNLP 2024, 2024

    2024

  31. [32]

    Crimson: A clinically-grounded llm-based metric for generative radiology report evaluation.arXiv preprint arXiv:2603.06183, 2026

    Mohammed Baharoon, Thibault Heintz, Siavash Raissi, Mahmoud Alabbad, Mona Alham- mad, Hassan AlOmaish, Sung Eun Kim, Oishi Banerjee, and Pranav Rajpurkar. Crimson: A clinically-grounded llm-based metric for generative radiology report evaluation.arXiv preprint arXiv:2603.06183, 2026

    arXiv 2026

  32. [33]

    Lungren, Andrew Y

    Saahil Jain, Ashwin Agrawal, Adriel Saporta, Steven Truong, Du Nguyen Duong, Tan Bui, Pierre Chambon, Yuhao Zhang, Matthew P. Lungren, Andrew Y . Ng, Curtis Langlotz, and Pranav Rajpurkar. Radgraph: Extracting clinical entities and relations from radiology reports. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks ...

    2021

  33. [34]

    Radeval: A framework for radiology text evaluation

    Justin Xu, Xi Zhang, Javid Abderezaei, Julie Bauml, Roger Boodoo, Fatemeh Haghighi, Ali Ganjizadeh, Eric Brattain, Dave Van Veen, Zaiqiao Meng, et al. Radeval: A framework for radiology text evaluation. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 2025

    2025

  34. [35]

    Report the imaging findings in the liver and biliary tree for this abdominal CT scan

    Yu Xin, Gorkem Can Ates, Kuang Gong, and Wei Shao. Med3dvlm: An efficient vision- language model for 3d medical image analysis.IEEE Journal of Biomedical and Health Informatics, 2025. 12 Appendix A Dataset Statistics Table 6:Dataset statistics.CT volumes per dataset, train / test split, and finding taxonomy size. ⋆Private out-of-distribution evaluation se...

    arXiv 2025