Pith. sign in

REVIEW 4 major objections 4 minor 32 references

Beyond Pixel Agreement: Large Language Models as Clinical Guardrails for Reliable Medical Image Segmentation

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a large language model driven by a structured multi-stage prompt can act as a zero-shot clinical guardrail for AI-generated medical image segmentations, matching or exceeding the accuracy of vision classifiers…

desk verdict The LLM-as-QC idea is new and worth testing, but the reported numbers can't all come from the same test set, so the main comparison needs raw data before it can be trusted. read the letter →

arxiv 2506.01841 v1 pith:PRN3SSP3 submitted 2025-06-02 eess.IV

Jiaxi Sheng , Leyi Yu , Haoyue Li , Yifan Gao , Xin Gao This is my paper · ORCID
classification eess.IV
keywords LargeLanguageModelsMedicalImageSegmentationAutomaticQualityControlHierarchicalClinicalReasonerZero-shotevaluationacceptabilityInterpretablereasoningGemini2.5Flash
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 introduces the Hierarchical Clinical Reasoner (HCR), a framework that uses a large language model as a zero-shot quality inspector for AI-generated medical image segmentations. Its central claim is that, guided by a structured prompt that walks the model through knowledge recall, visual feature analysis, anatomical inference, and clinical synthesis, an LLM can classify segmentations as clinically acceptable or not at 78.12% accuracy, matching or beating a ResNet50 trained for the task (72.92%). If true, this would mean clinical quality control for segmentation no longer needs task-specific training data for a dedicated classifier, and the step-by-step reasoning the LLM emits could give clinicians a transparent audit of why a case was accepted or rejected.

What carries the argument

The load-bearing mechanism is the Hierarchical Clinical Reasoner prompt, a hand-authored structured text prompt that instructs the LLM to first recall modality-specific anatomy and imaging characteristics, then analyze simple visual features of the segmentation contour such as continuity, edge contrast, and internal texture, then infer anatomical plausibility by checking for under-segmentation and spillage, and finally synthesize a clinical conclusion with a numerical score and an accept or reject recommendation. The prompt converts an otherwise unguided visual question into a forced, stepwise clinical reasoning trace that is reduced to a binary label by a fixed rule.

What would settle it

Have at least three independent clinical experts re-label the same 96 test images using the paper's 1 to 5 usability scale and compute inter-rater agreement, such as Fleiss' kappa; if agreement falls below roughly 0.6, the original single-panel ground truth is too noisy to support the claimed 78.12% accuracy or the comparison to trained baselines.

Watch

Extended reading notes

Core claim

On the paper's own terms, HCR powered by Gemini 2.5 Flash operates zero-shot on a 96-case test set spanning six imaging tasks and modalities and classifies nnU-Net segmentations as accept or reject with 78.12% accuracy and an F1-score of 0.8264, exceeding the best trained vision baseline (ResNet50 at 72.92% accuracy; EfficientNet-B0 at 71.88%; ViT-Base at 69.79%). The framework outputs structured reasoning across four stages—knowledge recall, low-level visual feature analysis, anatomical inference, and clinical synthesis—ending in a 1-5 usability score and a binary recommendation, providing interpretability that the trained classifiers do not offer.

Load-bearing premise

The ground truth accept or reject labels are assigned by clinical experts, but the paper does not report how many experts labeled each case, their specialties, or inter-rater agreement, so the accuracy numbers for both HCR and the baselines are only as meaningful as those expert labels.

Editorial extensions

If this is right

  • A zero-shot LLM can serve as a clinical guardrail for segmentation quality control without requiring task-specific training labels for the evaluator.
  • The structured reasoning output gives an interpretable audit trail, in contrast to black-box pixel-metric or classifier-based QC methods.
  • HCR can be applied to a new segmentation task by writing a new prompt rather than training a new model.
  • Quality assessment built on clinical reasoning rather than pixel overlap can capture clinically relevant errors, such as small omissions or spillage into adjacent structures, that conventional metrics may miss.
  • Performance depends strongly on the underlying LLM, so choosing or ensembling stronger models could further improve QC accuracy.

Reading between the lines

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

  • If the LLM's reasoning is faithful, a natural deployment would be a two-stage pipeline: a cheap pixel-metric filter flags suspicious cases, and HCR reviews only those, cutting LLM inference cost.
  • The accept and reject labels would ideally be validated against downstream clinical consequences, such as dosimetric impact in radiotherapy planning, rather than expert opinion alone; the paper's ground-truth setup leaves that connection untested.
  • The four-stage prompt structure is modality-agnostic and may transfer to other clinical image-interpretation tasks, including quality assessment of 3D segmentations or registration outputs.
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 / 4 minor

Summary. The paper introduces Hierarchical Clinical Reasoner (HCR), a structured multi-stage prompting framework that asks a large language model to evaluate AI-generated medical image segmentations and issue a clinical usability recommendation. The authors curate 479 image-label pairs from a public dataset across six organ/modality groups, split them 80/20, train three vision baselines on the training split, and run four LLMs zero-shot through HCR on the test split. The central claim is that HCR with Gemini 2.5 Flash reaches 78.12% accuracy, outperforming trained baselines such as ResNet50 at 72.92%, while additionally generating interpretable rationales. The paper frames this as evidence that zero-shot LLMs, when guided by structured reasoning, can serve as clinical guardrails for segmentation quality control.

Significance. If the central claim were fully supported, this would be a useful contribution: a zero-shot, interpretable quality gate that does not require per-task training would have practical value in medical imaging pipelines, and the paper names this possibility explicitly. The use of a public dataset, the inclusion of trained vision baselines, and the explicit attempt to move beyond pixel-level metrics are all strengths. The reported reasoning outputs (Figure 3) do illustrate an interpretability advantage over conventional classifiers. However, the quantitative evidence is currently not trustworthy: Table 2 is internally inconsistent under standard definitions of precision and recall, the headline improvement over ResNet50 is not tested for statistical significance, and the ground-truth labeling procedure is not documented. These issues are load-bearing because every accuracy, precision, recall, and F1 number in the paper derives from the same 96-case test set.

major comments (4)
  1. [Table 2 / Sec. 4.2] For a binary classifier on N=96 cases, the underlying number of positive-class cases p is determined by accuracy A, precision P, and recall R through p = 96(1-A)/(1 - R(2 - 1/P)). Applying this identity to Table 2 gives p ≈ 39 for all three trained baseline rows and p ≈ 58 for all four HCR rows. Since all rows are evaluated on the same 96-case test set, these two groups of counts cannot both be correct; this inconsistency is not explained by label noise, because the same gold labels would have to yield two different class distributions. The authors must release per-case predictions, a single label file, and an explicit statement of which class (accept or reject) is treated as positive in each row before the headline 78.12% versus 72.92% comparison can be interpreted.
  2. [Sec. 4.1 / Table 2] The reported difference between HCR with Gemini 2.5 Flash (78.12%) and ResNet50 (72.92%) corresponds to roughly 5 correct classifications out of 96. No confidence intervals, bootstrap estimates, or significance tests are reported, and the evaluation uses a single 80/20 split. Under these conditions, the claim that HCR performs 'comparably to, and in instances exceeding' the trained baselines is not statistically supported. Please report interval estimates and a paired test such as McNemar's test, and also report the accept/reject class distribution in the test set.
  3. [Sec. 4.1 / Sec. 3.3] The paper states that the final quality classification is 'derived based on the synthesized score and predefined rules,' but the actual mapping from the 1-5 score to accept/reject is never specified. Without this mapping, the reported accuracy numbers cannot be reproduced, and the LLM's score distribution cannot be audited. Please provide the exact threshold(s), the handling of non-conforming outputs, and the full prompt text (e.g., in an appendix), since HCR is defined by that prompt.
  4. [Sec. 3.2] The ground-truth labels are described only as being assigned by 'clinical experts,' with no information about the number of reviewers per case, their specialties, blinding to the AI-generated nature of the segmentations, or inter-rater agreement. Because every metric in Table 2 is computed against these labels, noisy or biased expert labels would make all accuracy numbers uninterpretable. Please report the annotation procedure, the number of raters, and an agreement measure such as Cohen's kappa or a documented adjudication rule.
minor comments (4)
  1. [Sec. 3.2 / Table 1] The dataset is said to originate from reference [28], but the paper does not specify the exact version or preprocessing details for slice selection, windowing, or overlay colors; please add these details for reproducibility.
  2. [Sec. 2] The phrase 'To our best knowledge' should be 'To the best of our knowledge.'
  3. [Figures 1 and 2] The text and labels in Figures 1 and 2 are very small and will likely be illegible in the printed version; higher-resolution or vector versions are needed.
  4. [Sec. 5] The Limitations paragraph acknowledges that LLMs 'can occasionally produce inconsistent or unfaithful reasoning,' but the paper does not quantify how often this occurred or how such cases were handled in the reported accuracy; please add a small analysis of output-format compliance and reasoning consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: HCR is a zero-shot LLM benchmark against externally assigned expert labels; no fitted parameter is renamed as a prediction.

full rationale

The paper's central claim is that the zero-shot HCR prompt achieves accuracy comparable to trained vision classifiers. The HCR pipeline (Section 3.3) fits no parameters to the test set; Section 4.1 explicitly states that 'the HCR framework operated in a zero-shot manner on the test set.' The ground-truth labels were assigned by clinical experts (Section 3.2) on the basis of clinical usability, and the LLM's scores are empirical predictions against those labels. The fact that the HCR prompt uses the same 1-5 usability rubric that defines the accept/reject labels is task specification, not a derivation of the conclusion from its inputs: the LLM must still infer the correct score from each image. Self-citations in the reference list concern prior segmentation methods and are not load-bearing for the QC evaluation result. The internal inconsistency of Table 2 (baseline rows and HCR rows implying different numbers of positive cases on the same 96-case test set) is a serious reproducibility or correctness concern, but it is not circularity: even if the table is wrong, no claim is true by construction. Hence no circular step is exhibited, and the score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

HCR is a prompt-based framework and introduces no new physical or mathematical entities. The quantitative claims rest on hand-chosen thresholds and several domain assumptions about expert labels, 2D slices, and LLM visual perception.

free parameters (2)
  • HCR score-to-class threshold = Not disclosed (implied score >= 4 maps to accept)
    Section 4.1 states the final classification is derived from the synthesized score and predefined rules, but the exact threshold is never given. This hand-chosen boundary directly changes the reported accuracy.
  • Clinical usability label threshold = Expert ratings 4-5 -> accept, 1-3 -> reject
    Section 3.2 maps the 5-point expert scale to binary labels. This threshold determines the class balance and every downstream metric, and is set by the authors without sensitivity analysis.
assumptions (4)
  • domain assumption Clinical expert ratings on a 5-point usability scale are reliable and consistent enough to serve as ground truth.
    Invoked in Section 3.2 when assigning accept/reject labels; no inter-rater reliability, number of experts, or rating protocol is reported.
  • domain assumption A single 2D slice of a 3D medical scan contains sufficient information to judge the clinical acceptability of a segmentation.
    The curated dataset uses representative 2D slices (Section 3.2) and both experts and HCR evaluate only these slices; clinical acceptability can depend on volumetric context.
  • domain assumption The multimodal LLM perceives the overlaid segmentation contour accurately from the input image.
    HCR asks the LLM to analyze visual features of the overlay (Section 3.3), but the paper does not verify the image resolution, overlay color choices, or whether the model sees the contour reliably.
  • domain assumption The structured reasoning stages in the prompt elicit faithful and useful reasoning from the LLM.
    The Discussion (Section 5) admits LLMs can produce inconsistent or unfaithful reasoning, yet the interpretability claims of HCR depend on this reasoning being meaningful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Pixel Agreement: Large Language Models as Clinical Guardrails for Reliable Medical Image Segmentation." pith.science (2026). https://pith.science/paper/PRN3SSP3

@misc{pith2026250601841,
  author       = {Pith},
  title        = {Pith review of: Beyond Pixel Agreement: Large Language Models as Clinical Guardrails for Reliable Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PRN3SSP3}},
  note         = {Machine review of arXiv:2506.01841}
}
read the original abstract

Evaluating AI-generated medical image segmentations for clinical acceptability poses a significant challenge, as traditional pixelagreement metrics often fail to capture true diagnostic utility. This paper introduces Hierarchical Clinical Reasoner (HCR), a novel framework that leverages Large Language Models (LLMs) as clinical guardrails for reliable, zero-shot quality assessment. HCR employs a structured, multistage prompting strategy that guides LLMs through a detailed reasoning process, encompassing knowledge recall, visual feature analysis, anatomical inference, and clinical synthesis, to evaluate segmentations. We evaluated HCR on a diverse dataset across six medical imaging tasks. Our results show that HCR, utilizing models like Gemini 2.5 Flash, achieved a classification accuracy of 78.12%, performing comparably to, and in instances exceeding, dedicated vision models such as ResNet50 (72.92% accuracy) that were specifically trained for this task. The HCR framework not only provides accurate quality classifications but also generates interpretable, step-by-step reasoning for its assessments. This work demonstrates the potential of LLMs, when appropriately guided, to serve as sophisticated evaluators, offering a pathway towards more trustworthy and clinically-aligned quality control for AI in medical imaging.

Figures

Figures reproduced from arXiv: 2506.01841 by the authors.

Figure 1
Figure 1. Conceptual overview of the medical image segmentation quality assessment task addressed in this work. The illustration outlines (a) the core components presented for evaluation, including task instructions, an image sample, and a guiding question; (b) the traditional clinician-centric evaluation process; and (c) our proposed LLM-driven evaluation process facilitated by the Hierarchical Clinical Reasoner (HCR). HCR f… view at source ↗
Figure 2
Figure 2. An overview of the Hierarchical Clinical Reasoner (HCR) framework, illustrat￾ing the multi-stage process from input (medical image with AI segmentation) to the structured clinical quality assessment generated by the LLM. 3.1 Method Overview The goal of our Hierarchical Clinical Reasoner (HCR) is to provide a clinically￾aligned quality assessment of AI-generated medical image segmentations using [PITH_FULL_IMAGE:fig… view at source ↗
Figure 3
Figure 3. Illustrative example of the detailed clinical reasoning and structured output generated by different LLMs within the HCR framework for representative segmenta￾tion cases. These examples highlight the interpretability afforded by our approach. 4.3 Qualitative Analysis and Interpretability Beyond quantitative metrics, a key attribute of our HCR approach is its ability to generate detailed, interpretable textual explan… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 22 canonical work pages

  1. [1]

    Wega: Weakly-supervised global-local affinity learning framework for lymph node metastasis prediction in rectal cancer

    Yifan Gao, Yaoxian Dong, Wenbin Wu, Chaoyang Ge, Feng Yuan, Jiaxi Sheng, Haoyue Li, and Xin Gao. Wega: Weakly-supervised global-local affinity learning framework for lymph node metastasis prediction in rectal cancer. arXiv preprint arXiv:2505.10502, 2025

    arXiv 2025

  2. [2]

    Multi-modal disease segmentation with continual learning and adaptive decision fusion

    Xu Xu, Junxin Chen, Dipanwita Thakur, and Duo Hong. Multi-modal disease segmentation with continual learning and adaptive decision fusion. Information Fusion, page 102962, 2025

    2025

  3. [3]

    Tooth instance segmentation and disease detection with uncertainty-aware contrastive learning and cross-scale attention

    Xu Xu, Junxin Chen, and Jiayue Yin. Tooth instance segmentation and disease detection with uncertainty-aware contrastive learning and cross-scale attention. IEEE Journal of Biomedical and Health Informatics , 2025

    work page 2025

  4. [4]

    Transmed: Transformers advance multi-modal medical image classification

    Yin Dai, Yifan Gao, and Fayu Liu. Transmed: Transformers advance multi-modal medical image classification. Diagnostics, 11(8):1384, 2021

    2021

  5. [5]

    Facial video-based non- contact stress recognition utilizing multi-task learning with peak attention

    Juncong Xu, Cheng Song, Zijie Yue, and Shuai Ding. Facial video-based non- contact stress recognition utilizing multi-task learning with peak attention. IEEE Journal of Biomedical and Health Informatics , 2024

    work page 2024

  6. [6]

    An anatomy-aware framework for automatic segmentation of parotid tumor from multimodal mri

    Yifan Gao, Yin Dai, Fayu Liu, Weibing Chen, and Lifu Shi. An anatomy-aware framework for automatic segmentation of parotid tumor from multimodal mri. Computers in Biology and Medicine , 161:107000, 2023

    work page 2023

  7. [7]

    Desam: Decoupled segment anything model for generalizable medical image segmentation

    Yifan Gao, Wei Xia, Dingdu Hu, Wenkui Wang, and Xin Gao. Desam: Decoupled segment anything model for generalizable medical image segmentation. In Inter- national Conference on Medical Image Computing and Computer-Assisted Inter- vention, pages 509–519. Springer, 2024

    2024

  8. [8]

    Mba-net: Sam-driven bidirec- tional aggregation network for ovarian tumor segmentation

    Yifan Gao, Wei Xia, Wenkui Wang, and Xin Gao. Mba-net: Sam-driven bidirec- tional aggregation network for ovarian tumor segmentation. In International Con- ference on Medical Image Computing and Computer-Assisted Intervention , pages 437–447. Springer, 2024

    work page 2024

Show all 32 references
  1. [9]

    Abs-mamba: Sam2-driven bidirectional spiral mamba network for medical image translation

    Feng Yuan, Yifan Gao, Wenbin Wu, Keqing Wu, Xiaotong Guo, Jie Jiang, and Xin Gao. Abs-mamba: Sam2-driven bidirectional spiral mamba network for medical image translation. arXiv preprint arXiv:2505.07687 , 2025

    2025

  2. [10]

    Medical image segmentation review: The suc- cess of u-net

    Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Had- dadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The suc- cess of u-net. IEEE Transactions on Pattern Analysis ...

    2024

  3. [11]

    Qualitative evaluation of common quantitative metrics for clinical acceptance of automatic segmentation: a case study on heart contouring from ct images by deep learning algorithms

    LB Van Den Oever, W A Van Veldhuizen, Ludo J Cornelissen, DS Spoor, Tineke P Willems, G Kramer, T Stigter, Mieneke Rook, APG Crijns, Matthijs Oudkerk, et al. Qualitative evaluation of common quantitative metrics for clinical acceptance of automatic segmentation: a case study o...

    2022

  4. [12]

    No-reference seg- mentation annotation quality assessment

    Zheng Lin, Zheng-Peng Duan, Xuying Zhang, and Luojun Lin. No-reference seg- mentation annotation quality assessment. In 2024 IEEE International Conference on Multimedia and Expo (ICME) , pages 1–6. IEEE, 2024

    2024

  5. [13]

    Medical image segmentation automatic quality control: A multi- dimensional approach

    Joris Fournel, Axel Bartoli, David Bendahan, Maxime Guye, Monique Bernard, Elisa Rauseo, Mohammed Y Khanji, Steffen E Petersen, Alexis Jacquier, and Badih Ghattas. Medical image segmentation automatic quality control: A multi- dimensional approach. Medical Image Analysis, 74:1...

    2021

  6. [14]

    Sqa-sam: seg- mentation quality assessment for medical images utilizing the segment anything model

    Yizhe Zhang, Shuo Wang, Tao Zhou, Qi Dou, and Danny Z Chen. Sqa-sam: seg- mentation quality assessment for medical images utilizing the segment anything model. arXiv preprint arXiv:2312.09899 , 2023. Hierarchical Clinical Reasoner 11

    2023 arXiv

  7. [15]

    A robust quality estimation method for medical image segmentation with small datasets

    Fatmat¨ ulzehra Uslu and Marta Varela. A robust quality estimation method for medical image segmentation with small datasets. Biomedical Signal Processing and Control, 95:106300, 2024

    2024

  8. [16]

    Segmentation quality assessment by automated detection of erroneous surface regions in medical images

    Fahim Ahmed Zaman, Lichun Zhang, Honghai Zhang, Milan Sonka, and Xiaodong Wu. Segmentation quality assessment by automated detection of erroneous surface regions in medical images. Computers in biology and medicine , 164:107324, 2023

    2023

  9. [17]

    Trustworthy clinical ai solutions: A unified review of uncertainty quantifi- cation in deep learning models for medical image analysis

    Benjamin Lambert, Florence Forbes, Senan Doyle, Harmonie Dehaene, and Michel Dojat. Trustworthy clinical ai solutions: A unified review of uncertainty quantifi- cation in deep learning models for medical image analysis. Artificial Intelligence in Medicine, 150:102830, 2024

    2024

  10. [18]

    Efficient bayesian uncertainty estimation for nnu-net

    Yidong Zhao, Changchun Yang, Artur Schweidtmann, and Qian Tao. Efficient bayesian uncertainty estimation for nnu-net. In International Conference on Medical Image Computing and Computer-Assisted Intervention , pages 535–544. Springer, 2022

    2022

  11. [19]

    Quality sentinel: Estimat- ing label quality and errors in medical segmentation datasets

    Yixiong Chen, Zongwei Zhou, and Alan Yuille. Quality sentinel: Estimat- ing label quality and errors in medical segmentation datasets. arXiv preprint arXiv:2406.00327, 2024

    2024 arXiv

  12. [20]

    A survey on evaluation of large language models

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM transactions on intelligent systems and technology , 15(3):1–45, 2024

    2024

  13. [21]

    Large language models in medicine

    Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930–1940, 2023

    1930

  14. [22]

    Application of large language models in medicine

    Fenglin Liu, Hongjian Zhou, Boyang Gu, Xinyu Zou, Jinfa Huang, Jinge Wu, Yiru Li, Sam S Chen, Yining Hua, Peilin Zhou, et al. Application of large language models in medicine. Nature Reviews Bioengineering, pages 1–20, 2025

    2025

  15. [23]

    Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation

    Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22227–2...

    2024

  16. [24]

    Multimodal large language models address clinical queries in laryngeal cancer surgery: a com- parative evaluation of image interpretation across different models

    Bingyu Liang, Yifan Gao, Taibao Wang, Lei Zhang, and Qin Wang. Multimodal large language models address clinical queries in laryngeal cancer surgery: a com- parative evaluation of image interpretation across different models. International Journal of Surgery , 111(3):2727–2730, 2025

    2025

  17. [25]

    Bliva: A simple multimodal llm for better handling of text-rich visual questions

    Wenbo Hu, Yifan Xu, Yi Li, Weiyue Li, Zeyuan Chen, and Zhuowen Tu. Bliva: A simple multimodal llm for better handling of text-rich visual questions. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 2256–2264, 2024

    2024

  18. [26]

    Large language models are latent variable models: Explaining and find- ing good demonstrations for in-context learning

    Xinyi Wang, Wanrong Zhu, Michael Saxon, Mark Steyvers, and William Yang Wang. Large language models are latent variable models: Explaining and find- ing good demonstrations for in-context learning. Advances in Neural Information Processing Systems, 36:15614–15638, 2023

    2023

  19. [27]

    Prompt engineering in consistency and reliability with the evidence- based guideline for llms

    Li Wang, Xi Chen, XiangWen Deng, Hao Wen, MingKe You, WeiZhi Liu, Qi Li, and Jian Li. Prompt engineering in consistency and reliability with the evidence- based guideline for llms. NPJ digital medicine , 7(1):41, 2024

    2024

  20. [28]

    Ai-generated annotations dataset for diverse cancer radiol- ogy collections in nci image data commons

    Gowtham Krishnan Murugesan, Diana McCrumb, Mariam Aboian, Tej Verma, Rahul Soni, Fatima Memon, Keyvan Farahani, Linmin Pei, Ulrike Wagner, An- drey Y Fedorov, et al. Ai-generated annotations dataset for diverse cancer radiol- ogy collections in nci image data commons. Scientif...

    2024

  21. [29]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation

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

    2021

  22. [30]

    Efficientnet: Rethinking model scaling for convolu- tional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolu- tional neural networks. In International conference on machine learning , pages 6105–6114. PMLR, 2019

    2019

  23. [31]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  24. [32]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International C...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.