REVIEW 5 major objections 5 minor 35 references
Look & Mark: Leveraging Radiologist Eye Fixations and Bounding boxes in Multimodal Large Language Models for Chest X-ray Report Generation
T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper claims that a training-free prompting strategy combining radiologist eye-fixation durations with abnormality bounding boxes improves chest X-ray report generation across multiple multimodal language models and, in expert…
desk verdict A prompt recipe that combines gaze and boxes for CXR reports, but the oracle-assisted evaluation overstates consistent gains. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the grounding fixation prompt, which has two channels. A visual channel renders abnormality bounding boxes $B=\{b_1,\ldots,b_n\}$ directly on the image; a textual channel encodes fixation data $G=\{g_1,\ldots,g_m\}$ as per-box durations. The load-bearing identity is the mapping $M(g_j,B)=b_i$ when fixation point $(x_j,y_j)$ lies inside box $b_i$, with ties resolved to the smallest box, and the summed duration $T_i=\sum_{g_j \in G,\,M(g_j,B)=b_i} t_j$. This converts raw eye-tracking output into a short prompt stating which abnormality received how much looking time, while the image carries the spatial marks. In-context learning with three dictated-report exemplars is the mechanism that lets general-purpose models adopt a clinical writing style without weight updates; the paper claims these pieces together produce grounded, hallucination-reduced reports.
What would settle it
Run the same Look & Mark prompts on a held-out test set where the bounding boxes and fixation durations are produced by an automated detector and fixation predictor instead of taken from the auxiliary datasets; if the improvement over baseline prompting collapses toward zero, the reported gains depend on oracle labels at inference time rather than on the prompt strategy itself.
Extended reading notes
Core claim
On the paper's own terms, the discovery is that expert spatial and temporal knowledge can be injected as prompts rather than learned. Overlaying abnormality bounding boxes on the X-ray and listing per-abnormality fixation durations in the text steers a frozen multimodal LLM toward clinically grounded reports. The paper formalizes fixations as points with durations, maps each fixation to the smallest enclosing bounding box, sums the fixation time per box, and formats the result as 'Fixation Data: [Abnormality bounding box: {label}, Fixation Time: {time} seconds]'. It reports that this combined prompt beats either cue alone on most models, and that adding three in-context report exemplars lets general-purpose models match or beat domain-specific report generators; the one clear exception is MAIRA2, a model already trained for grounded report generation, which shows no gain from the extra fixation cue.
Load-bearing premise
The method assumes that ground-truth abnormality bounding boxes and eye-fixation data are available at inference time for every test image, so the measured gains may reflect oracle information rather than a generally deployable prompt strategy; the paper's limitations section acknowledges the expert-input requirement.
Editorial extensions
If this is right
- Any multimodal LLM, including one never trained on chest X-rays, can be steered toward CXR report generation with the same prompt, making the method a drop-in alternative to fine-tuning.
- Combining Look and Mark beats either cue alone on most models, and the Mark (bounding-box) channel generally contributes more than the Look (fixation) channel, so spatial grounding is the stronger of the two signals.
- A general-purpose model, LLaVA-OV with in-context learning and L&M, reaches the highest clinical average (87.3% C.AVG) of all tested configurations, above models explicitly trained for CXR report generation.
- Three radiologists scored CXR-LLaVA with L&M at 1.75 clinically significant errors per report versus 2.18 for its baseline prompt, a reduction of 0.43 errors per report.
- The largest gains appear in clinical relevance metrics (RaTEScore and RadGraph-XL) rather than lexical overlap, consistent with the claim that grounding reduces false and omitted findings more than it improves wording.
Reading between the lines
- The reported numbers are an oracle-assisted ceiling: the test-time prompt includes the ground-truth bounding boxes and fixation durations, so a deployed system would need automated abnormality detection and fixation prediction to reproduce them; the paper's limitations section lists this reliance on expert input and leaves automated grounding for future work.
- If fixation durations could be captured live with an eye tracker during a radiologist's read, the method becomes a human-in-the-loop aid that needs no stored annotations, an extension the paper does not test.
- The observation that Qwen2.5VL, a grounding-trained model, does not benefit from added fixation cues when Mark is present suggests that models already strong at spatial grounding may saturate, so the Look channel likely helps most when spatial grounding is weak.
- Because the gains concentrate in clinical metrics and only chest X-ray is tested, a natural next step is to see whether the same prompt structure transfers to other image-report domains such as CT or MRI.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes Look & Mark (L&M), a prompt-only strategy for chest X-ray report generation that injects radiologist eye-fixation durations and abnormality bounding boxes into the prompts of multimodal large language models. The method is evaluated on 560 images from REFLACX/MS_CXR with six models, comparing default prompting, fixation-only, bounding-box-only, and combined L&M prompting, with and without in-context learning. The authors report improvements in lexical and clinical metrics, claim consistent gains across models, and support this with a small radiologist error-annotation study. The underlying idea is simple and potentially useful, but the evaluation as presented does not support the central claims because the test-time prompts contain ground-truth abnormality labels, and several reported results contradict the 'consistent gains' narrative.
Significance. If the claimed gains were obtained in a deployment-realistic setting, L&M would be a valuable training-free contribution: combining spatial marks and temporal gaze cues in a prompt is a plausible way to reduce hallucinations without fine-tuning. The paper has strengths: it uses publicly available REFLACX and MS_CXR data, evaluates a wider range of domain-specific and general-purpose models than many prior works, and includes an expert evaluation with reported inter-annotator agreement. However, the significance of these results is currently undermined by the oracle nature of the test-time inputs, the absence of significance testing, and internal inconsistencies in the reported numbers. As an oracle upper-bound study, the work may still be of interest, but the abstract and conclusion present it as a practical, scalable solution, which the evidence does not support.
major comments (5)
- [Abstract; Section 4.1; Section 3.2; Eqs. (1)-(4)] The evaluation setting gives the model the answer it is supposed to produce. Each test image is annotated with MS_CXR ground-truth bounding boxes, and Eq. (3) maps REFLACX fixation points to those boxes, yielding Eq. (4) fixation times per abnormality label. Section 3.2 then formats this as 'Abnormality bounding box: {label}, Fixation Time: {time} seconds' in the prompt. Thus the L&M prompt explicitly states the abnormality labels that the report is asked to generate. The comparison between '-' and L&M therefore measures label leakage as well as grounding, and the reported gains are an oracle upper bound, not evidence for a deployment-ready method. The Limitations section mentions automated grounding only as future work, so the authors do not claim the current system is deployable, but the abstract's statement that L&M is a 'scalable and efficient solution' and the conclusion's 'practical solution for low-resource clinical environments' are not supported by the evaluated setting.
- [Section 5.1; Table 4; Table 6] The abstract claims 'consistent gains' across models, but Table 4 shows multiple decreases. MAIRA2 L&M lowers C.AVG from 74.35 to 73.16 and A.AVG from 66.70 to 66.31 relative to the default prompt. Qwen2.5VL L&M lowers A.AVG from 50.08 to 48.71 and also lowers ROUGE-L and BERTScore. In Appendix Table 6, Qwen2.5VL I&L&M is substantially worse than Qwen2.5VL I&M on every clinical metric, with C.AVG dropping from 88.53 to 74.83 and RadGraph-XL from 0.1175 to 0.0812. These results directly contradict the narrative of consistent improvement and need to be acknowledged and analyzed rather than explained away in passing.
- [Section 5.2.2; Table 6] The text in Section 5.2.2 states that for Qwen2.5VL, 'I&L&M outperforms I&L with a small incarse of 0.6%', but Table 6 shows the opposite: I&L&M has A.AVG 55.88 versus I&L 58.38, a decrease of 2.50 percentage points, and C.AVG 74.83 versus 80.02, a decrease of 5.19 percentage points. Similarly, the claim that 'I&L&M leads to significant performance gains over I&L in all models' is false for Qwen2.5VL on both A.AVG and C.AVG. This is a factual error in the results discussion that must be corrected. Additionally, in Table 6 the MAIRA2 M and MAIRA2 L&M rows are numerically identical, which either means the fixation cue had no effect for this model or is a data transcription error; either way it requires clarification.
- [Section 4.3; Table 4; Table 6] The paper reports no confidence intervals, significance tests, or variance estimates for any of the reported metrics. Several headline improvements are small relative to likely noise, for example CXR-LLaVA A.AVG +1.19 percentage points and MAIRA2 ROUGE-L +0.0009. Without statistical testing or multiple-run variability, the words 'significant gains' throughout Section 5 are not justified. In addition, the definition of C.AVG and A.AVG as normalized by the highest observed score makes these values relative to the specific set of methods in the table; a score of 87.3% does not mean 87.3% of some absolute clinical quality. This normalization choice should be stated clearly whenever comparative claims such as 'highest among all models' are made.
- [Section 5.3; Table 5] The expert evaluation is central to the claim that L&M reduces hallucinations, but the reporting is incomplete. Table 5 gives average errors per report for only five model-method combinations, with no statement of how many reports were scored, how the reports were sampled, whether the same images were used across all methods, or whether the reported differences (e.g., 1.75 versus 2.18) are statistically significant. The Krippendorff alpha of 0.647 is useful, but it does not by itself support a 0.43-error reduction claim. This section needs substantially more detail before the hallucination-reduction claim can be evaluated.
minor comments (5)
- [Section 5.2.2] There is a typo: 'incarse' should be 'increase'. The surrounding sentence also needs to be rewritten because, as noted above, the direction of the Qwen2.5VL result is reversed.
- [Section 5.1.1] There is a typo: 'decresae' should be 'decrease'. Please proofread the manuscript for similar spelling errors.
- [Abstract and throughout] The notation 'A.A VG' and 'C.A VG' appears with an inconsistent space in the abstract and in several places in the tables; these should be written uniformly as A.AVG and C.AVG.
- [Section 3.2; Section 4.3; Figure 1] Please clarify exactly what is provided visually versus textually for the L, M, and L&M conditions. Section 3.2 says fixation durations are provided as text, Section 4.3 says L corresponds to eye fixation data represented as heatmaps, and Figure 1 suggests both heatmaps and textual summaries are used. This ambiguity makes it difficult to reproduce the method.
- [Appendix; Figure 4] The appendix says the three examples in Figure 4 can also be regarded as the three exemplar reports used for in-context learning. If these are the same examples, this should be stated explicitly in the main text so that readers can assess whether the qualitative examples are independent of the ICL exemplars.
Circularity Check
Oracle-assisted prompting: the prompt injects ground-truth abnormality labels and gaze totals (Eqs. 1, 3-4, Section 3.2), so the reported clinical-metric gains and expert error reductions largely reduce to restating the supplied labels; automated grounding is deferred to future work.
-
other
[Section 3.2 (Eqs. 1-4, prompt template), Section 4.1, Section 5.1/Table 4, Section 5.3/Table 5]
"The fixation data is formatted into textual prompts following this template: Fixation Data: [Abnormality bounding box: {label}, Fixation Time: {time} seconds]. ... we used two primary datasets: the REFLACX dataset as the source of eye fixation data and dictated reports, and the MS_CXR dataset as the source of abnormalities bounding boxes"
At inference, every test image is accompanied by MS_CXR ground-truth abnormality labels (Eq. 1: each bi carries label li) and gaze durations accumulated inside those same ground-truth boxes (Eqs. 3-4). The prompt spells the label out ('Abnormality bounding box: {label}, Fixation Time: {time} seconds'). The evaluation then rewards exactly those entities: RadGraph-XL/RaTEScore credit correct mentions, and the expert categories 'false prediction'/'omission' (5.3) are mechanically improved when the correct finding is already in the prompt.
full rationale
The main result is not formally circular in the fitted-parameter sense: no parameter is fitted from the test set, and the self-citations (Kim et al. 2024a,b, sharing authors with this paper) only motivate the use of gaze; the empirical ablation is self-contained and would stand on its own as a conditioning study. The serious issue is test-time conditioning on ground truth. Section 4.1 states that MS_CXR supplies the abnormality bounding boxes (Eq. 1) and REFLACX supplies the fixations and dictated reports; Section 3.2 maps every fixation to the ground-truth box it falls in (Eqs. 3-4) and then puts the ground-truth label and its fixation time verbatim into the prompt. The generated report is therefore produced with the abnormality label visible, i.e., the answer key is an input. The evaluation then rewards exactly that answer key: RadGraph-XL and RaTEScore give credit for mentioning the named entities, and the expert error categories (false prediction, omission) are mechanically improved when the label is supplied. The reported gains, such as LLaVA-Med RaTEScore jumping from 0.2445 to 0.4191 and C.AVG from 24.99% to 52.46%, and the expert-error reduction from 2.18 to 1.75 for CXR-LLaVA, largely measure the model's ability to restate the injected labels, not an emergent grounding-fixation skill; the baseline closed-book models simply failed to name those entities, and the prompt hands them the names. The paper is partially transparent about this: the Limitations section says expert-labeled boxes and fixations are required and that automatic fixation prediction and bounding box grounding is future work, and the Conclusion repeats that automated grounding is future work. That honesty prevents a self-citation-chain verdict, but the Abstract and Conclusion still frame L&M as a scalable, efficient, low-resource solution, while the only tested setting requires the ground-truth labels at inference. Hence the central 'prediction', the clinically improved report, reduces in substantial part to its own inputs, giving partial circularity (score 6) rather than a fully definitional 8-10. The internal max-normalization used to define C.AVG and A.AVG also makes the 'highest clinical average' headline partly self-referential, but that is a minor metric artifact relative to the label-injection issue.
Assumptions & free parameters
assumptions (3)
- domain assumption Fixation points inside a bounding box indicate attention to that abnormality (Eq. 3).
- domain assumption MS_CXR bounding boxes and REFLACX eye-fixation data can be merged on the same MIMIC-CXR images with consistent coordinates.
- domain assumption Ground-truth abnormality boxes and fixation durations are available at test time.
Cite this review
Pith. "Pith review of Look & Mark: Leveraging Radiologist Eye Fixations and Bounding boxes in Multimodal Large Language Models for Chest X-ray Report Generation." pith.science (2026). https://pith.science/paper/6D4EG7NH
@misc{pith2026250522222,
author = {Pith},
title = {Pith review of: Look & Mark: Leveraging Radiologist Eye Fixations and Bounding boxes in Multimodal Large Language Models for Chest X-ray Report Generation},
year = {2026},
howpublished = {\url{https://pith.science/paper/6D4EG7NH}},
note = {Machine review of arXiv:2505.22222}
}
read the original abstract
Recent advancements in multimodal Large Language Models (LLMs) have significantly enhanced the automation of medical image analysis, particularly in generating radiology reports from chest X-rays (CXR). However, these models still suffer from hallucinations and clinically significant errors, limiting their reliability in real-world applications. In this study, we propose Look & Mark (L&M), a novel grounding fixation strategy that integrates radiologist eye fixations (Look) and bounding box annotations (Mark) into the LLM prompting framework. Unlike conventional fine-tuning, L&M leverages in-context learning to achieve substantial performance gains without retraining. When evaluated across multiple domain-specific and general-purpose models, L&M demonstrates significant gains, including a 1.2% improvement in overall metrics (A.AVG) for CXR-LLaVA compared to baseline prompting and a remarkable 9.2% boost for LLaVA-Med. General-purpose models also benefit from L&M combined with in-context learning, with LLaVA-OV achieving an 87.3% clinical average performance (C.AVG)-the highest among all models, even surpassing those explicitly trained for CXR report generation. Expert evaluations further confirm that L&M reduces clinically significant errors (by 0.43 average errors per report), such as false predictions and omissions, enhancing both accuracy and reliability. These findings highlight L&M's potential as a scalable and efficient solution for AI-assisted radiology, paving the way for improved diagnostic workflows in low-resource clinical settings.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[1]
Rawan AlSaad, Alaa Abd-Alrazaq, Sabri Boughorbel, Arfan Ahmed, Max-Antoine Renault, Rafat Damseh, and Javaid Sheikh. 2024. Multimodal large language models in health care: Applications, challenges, and future outlook. Journal of Medical Internet Research, 26:e59505
work page 2024
-
[2]
Shruthi Bannur, Kenza Bouzid, Daniel C Castro, Anton Schwaighofer, Sam Bond-Taylor, Maximilian Ilse, Fernando P \'e rez-Garc \' a, Valentina Salvatelli, Harshita Sharma, Felix Meissen, et al. 2024. Maira-2: Grounded radiology report generation. arXiv preprint arXiv:2406.04449
arXiv 2024
-
[3]
Ricardo Bigolin Lanfredi, Mingyuan Zhang, William F Auffermann, Jessica Chan, Phuong-Anh T Duong, Vivek Srikumar, Trafton Drew, Joyce D Schroeder, and Tolga Tasdizen. 2022. Reflacx, a dataset of reports and eye-tracking data for localization of abnormalities in chest x-rays. Scientific data, 9(1):350
2022
-
[4]
Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, et al. 2022. Making the most of text semantics to improve biomedical vision--language processing. In European conference on computer vision, pages 1--21. Springer
work page 2022
-
[5]
Francisco Maria Calisto, Carlos Santiago, Nuno Nunes, and Jacinto C Nascimento. 2022. Breastscreening-ai: Evaluating medical intelligent agents for human-ai interactions. Artificial Intelligence in Medicine, 127:102285
work page 2022
-
[6]
Jiawei Chen, Dingkang Yang, Tong Wu, Yue Jiang, Xiaolu Hou, Mingcheng Li, Shunli Wang, Dongling Xiao, Ke Li, and Lihua Zhang. 2024 a . Detecting and evaluating medical hallucinations in large vision language models. arXiv preprint arXiv:2406.10185
arXiv 2024
-
[7]
Zhihong Chen, Maya Varma, Jean-Benoit Delbrouck, Magdalini Paschali, Louis Blankemeier, Dave Van Veen, Jeya Maria Jose Valanarasu, Alaa Youssef, Joseph Paul Cohen, Eduardo Pontes Reis, et al. 2024 b . Chexagent: Towards a foundation model for chest x-ray interpretation. arXiv preprint arXiv:2401.12208
arXiv 2024
-
[8]
Jean-Benoit Delbrouck, Pierre Chambon, Zhihong Chen, Maya Varma, Andrew Johnston, Louis Blankemeier, Dave Van Veen, Tan Bui, Steven Truong, and Curtis Langlotz. 2024. Radgraph-xl: A large-scale expert-annotated dataset for entity and relation extraction from radiology reports. In Findings of the Association for Computational Linguistics ACL 2024, pages 12...
work page 2024
Show all 35 references
-
[9]
Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783
2024 arXiv
-
[10]
Stephanie L Hyland, Shruthi Bannur, Kenza Bouzid, Daniel C Castro, Mercy Ranjit, Anton Schwaighofer, Fernando P \'e rez-Garc \' a, Valentina Salvatelli, Shaury Srivastav, Anja Thieme, et al. 2023. Maira-1: A specialised large multimodal model for radiology report generation. a...
2023 arXiv
-
[11]
Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. 2019. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317
2019
-
[12]
Yunsoo Kim, Jinge Wu, Yusuf Abdulle, Yue Gao, and Honghan Wu. 2024 a . Enhancing human-computer interaction in chest x-ray analysis using vision and language model with eye gaze patterns. arXiv preprint arXiv:2404.02370
2024 arXiv
-
[13]
Yunsoo Kim, Jinge Wu, Yusuf Abdulle, Yue Gao, and Honghan Wu. 2024 b . Human-in-the-loop chest x-ray diagnosis: Enhancing large multimodal models with eye fixation inputs. In International Workshop on Trustworthy Artificial Intelligence for Healthcare, pages 66--80. Springer
2024
-
[14]
Klaus Krippendorff. 2011. Content Analysis: An Introduction to Its Methodology. Sage Publications
2011
-
[15]
Seowoo Lee, Jiwon Youn, Mansu Kim, and Soon Ho Yoon. 2023. Cxr-llava: Multimodal large language model for interpreting chest x-ray images. arXiv preprint arXiv:2310.18341
2023 arXiv
-
[16]
Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. 2024 a . Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326
2024 arXiv
-
[17]
Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. 2024 b . Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36
2024
-
[18]
Yingshu Li, Yunyi Liu, Zhanyu Wang, Xinyu Liang, Lingqiao Liu, Lei Wang, Leyang Cui, Zhaopeng Tu, Longyue Wang, and Luping Zhou. 2023. A comprehensive study of gpt-4v's multimodal capabilities in medical imaging. medRxiv, pages 2023--11
2023
-
[19]
Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74--81
2004
-
[20]
Bhavik N. Patel, Louis Rosenberg, Gregg Willcox, David Baltaxe, Mimi Lyons, Jeremy Irvin, Pranav Rajpurkar, Timothy Amrhein, Rajan Gupta, Safwan Halabi, Curtis Langlotz, Edward Lo, Joseph Mammarappallil, A. J. Mariano, Geoffrey Riley, Jayne Seekins, Luyao Shen, Evan Zucker, an...
2019 doi
-
[21]
Trong Thang Pham, Ngoc-Vuong Ho, Nhat-Tan Bui, Thinh Phan, Patel Brijesh, Donald Adjeroh, Gianfranco Doretto, Anh Nguyen, Carol C Wu, Hien Nguyen, et al. 2024. Fg-cxr: A radiologist-aligned gaze dataset for enhancing interpretability in chest x-ray report generation. In Procee...
2024
-
[22]
Khaled Saab, Tao Tu, Wei-Hung Weng, Ryutaro Tanno, David Stutz, Ellery Wulczyn, Fan Zhang, Tim Strother, Chunjong Park, Elahe Vedadi, et al. 2024. Capabilities of gemini models in medicine. arXiv preprint arXiv:2404.18416
2024 arXiv
-
[23]
Akshay Smit, Saahil Jain, Pranav Rajpurkar, Anuj Pareek, Andrew Y Ng, and Matthew P Lungren. 2020. Chexbert: combining automatic labelers and expert annotations for accurate radiology report labeling using bert. arXiv preprint arXiv:2004.09167
2020 arXiv
-
[24]
Qwen Team. 2025. https://qwenlm.github.io/blog/qwen2.5-vl/ Qwen2.5-vl
2025
-
[25]
Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2023 a . Towards generalist foundation model for radiology. arXiv preprint arXiv:2308.02463
2023 arXiv
-
[26]
Jinge Wu, Yunsoo Kim, Eva C Keller, Jamie Chow, Adam P Levine, Nikolas Pontikos, Zina Ibrahim, Paul Taylor, Michelle C Williams, and Honghan Wu. 2023 b . Exploring multimodal large language models for radiology report error-checking. arXiv preprint arXiv:2312.13103
2023 arXiv
-
[27]
Jinge Wu, Yunsoo Kim, Daqian Shi, David Cliffton, Fenglin Liu, and Honghan Wu. 2024 a . Slava-cxr: Small language and vision assistant for chest x-ray report automation. arXiv preprint arXiv:2409.13321
2024 arXiv
-
[28]
Jinge Wu, Yunsoo Kim, and Honghan Wu. 2024 b . Hallucination benchmark in medical visual question answering. In The Second Tiny Papers Track at ICLR 2024
2024
-
[29]
Hanguang Xiao, Feizhong Zhou, Xingyue Liu, Tianqi Liu, Zhipeng Li, Xin Liu, and Xiaoxuan Huang. 2024. A comprehensive survey of large language models and multimodal large language models in medicine. arXiv preprint arXiv:2405.08603
2024 arXiv
-
[30]
Nur Yildirim, Hannah Richardson, Maria T Wetscherek, Junaid Bajwa, Joseph Jacob, Mark A Pinnock, Stephen Harris, Daniel Coelho de Castro, Shruthi Bannur, Stephanie L Hyland, et al. 2024. Multimodal healthcare ai: Identifying and designing clinically relevant vision-language ap...
2024 arXiv
-
[31]
Feiyang Yu, Mark Endo, Rayan Krishnan, Ian Pan, Andy Tsai, Eduardo Pontes Reis, Eduardo Kaiser Ururahy Nunes Fonseca, Henrique Min Ho Lee, Zahra Shakeri Hossein Abad, Andrew Y Ng, et al. 2023. Evaluating progress in automatic chest x-ray radiology report generation. Patterns, 4(9)
2023
-
[32]
Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675
2019 arXiv
-
[33]
Weike Zhao, Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2024. Ratescore: A metric for radiology report generation. medRxiv, pages 2024--06
2024
-
[34]
URL: " 'urlintro :=
ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before...
-
[35]
write newline
" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.