Pith. sign in

REVIEW 3 major objections 6 minor 39 references

A 3D VLM trained on LLM-refined synthetic reports beats 2D and 3D baselines on brain-MRI text tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 01:39 UTC pith:PRAMR5NZ

load-bearing objection A useful dataset-generation pipeline and striking VQA results, but the report-generation evaluation is circular and the paper overclaims; major revision is needed. the 3 major comments →

arxiv 2607.14581 v1 pith:PRAMR5NZ submitted 2026-07-16 cs.AI cs.CV

Multi-LLM Collaborative MRI Report Generation for Visual Instruction Tuning in Brain Oncology

Sinyoung Ra , Jonghun Kim , Hyunjin Park This is my paper
classification cs.AI cs.CV
keywords 3D medical imagingvision-language modelreport generationvisual question answeringvisual instruction tuningbrain MRImulti-LLM collaborationglioma and meningioma
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to solve a data-scarcity problem: 3D MRI scans have almost no paired radiology text, so medical vision-language models are largely stuck in 2D. The authors synthesize reports and question-answer sets for brain-tumor MRI by feeding structured tumor information to an LLM, then have three independent LLMs review and vote on each report, regenerating rejected ones with feedback. They train a 3D VLM on this refined data using visual instruction tuning. They claim this VLM outperforms 2D and 3D baselines on report-generation similarity and closed-ended visual question answering. A sympathetic reader would care because it suggests a scalable path to 3D medical VLMs without clinician annotation.

Core claim

The paper's central claim is that a 3D VLM trained on multi-LLM-refined synthetic text can understand whole MRI volumes well enough to write coherent radiology reports and correctly answer multiple-choice questions about tumor type, scan modality, and tumor location. On its own generated evaluation set, report BERT-F1 reaches roughly 80 and 79 on the two datasets compared with at most 70 for the best baseline, and VQA accuracy reaches roughly 92 and 89 compared with at most 54 for baselines. The authors also show that reports become more semantically separated by tumor type after multi-LLM refinement, which they read as improved clarity.

What carries the argument

The load-bearing mechanism is 3D visual instruction tuning with a frozen language model: a pretrained VQ encoder compresses each 3D MRI into latent codes, a 3D convolution flattens them into image tokens, and a 3D perceiver projects those tokens into the LLM's embedding space while low-rank adaptation fine-tunes the LLM. The training data is manufactured by a two-stage LLM pipeline: one LLM writes reports from structured features such as disease, modality, location, volume, and sphericity, and three reviewer LLMs vote on each report; any report rejected by two or more reviewers goes back to the writer with feedback. The voting loop is what removes inaccuracies and hallucinations before text

Load-bearing premise

The load-bearing premise is that LLM-generated and multi-LLM-refined text is accurate enough to be both the training target and the evaluation reference; the authors themselves concede that this shared origin may introduce a self-consistency bias, so the reported margins could measure imitation of the generator rather than clinical reading skill.

What would settle it

Take real hospital brain MRI volumes with independently written radiologist reports and run the trained VLM on them; if its BERT-F1 against those human reports is no better than the 2D baselines, or if blinded radiologists can reliably distinguish the generated reports from human ones, the central claim is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A 3D VLM can extract clinically relevant attributes such as tumor type, modality, and location from full MRI volumes, not just from a single 2D slice.
  • Synthetic, multi-reviewer text can substitute for scarce expert-annotated radiology pairs when building training datasets for 3D medical VLMs.
  • The same generation-plus-voting recipe should transfer to other 3D modalities like CT or PET, where paired text is equally scarce.
  • If the claim holds, building a brain-oncology report generator no longer requires a clinician in the data-creation loop.
  • Report similarity and cluster-separability improve after multi-LLM refinement, evidence that the voting and feedback step removes noise from the synthetic corpus.

Where Pith is reading between the lines

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

  • A human-reader test the paper leaves untested: blind radiologists comparing generated reports with genuine clinical reports would show whether the prose is clinically plausible or merely template-like.
  • Because both training and evaluation text come from the same LLM-linked pipeline, the quantitative margins may partly measure self-consistency; an external human-written benchmark would be needed to confirm the gains transfer.
  • The refinement loop could be turned into an active-learning signal: reports accepted unanimously by all three reviewers are likely cleaner, and one could test whether a single strong reviewer suffices or whether reviewer diversity is essential.
  • The modeling stack is not tumor-specific, so applying the same pipeline to other 3D lesion types or longitudinal scans is a natural next step that the paper does not explore.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This paper addresses the lack of paired 3D MRI-text data by synthesizing radiology reports and VQA items from BraTS-derived structured labels using GPT-4o-mini, then refining them through a multi-LLM review loop with Claude 3.5 Sonnet, Gemini 1.5 Flash-8B, and DeepSeek-chat under majority-vote accept/feedback. The authors train a 3D VLM consisting of a pre-trained VQ-GAN encoder, a 3D Perceiver, and Vicuna-7B with LoRA, and evaluate it on BraTS2021-GLI and BraTS2023-MEN. They report large gains over 2D (LLaVA, LLaVA-Med, ChatGPT) and 3D (M3D) baselines on report generation metrics (BLEU, ROUGE, METEOR, BERT-F1) and multiple-choice VQA accuracy.

Significance. The contribution is potentially useful: a scalable, LLM-based route to synthesizing 3D image-text datasets, a multi-LLM refinement pipeline, and a complete VLM training framework. The architecture choices are standard and clearly described. However, the central empirical claim is not yet established. Report-generation scores are computed against references produced by the same generation/refinement pipeline, the model comparison is confounded by unequal input information, and no uncertainty quantification or human evaluation is provided. The authors acknowledge the self-consistency bias in the Discussion, but the abstract and contributions still assert superiority. Since no code or data are released, the dataset contribution is currently not reproducible. If the evaluation issues are addressed, the work could become a valuable empirical contribution.

major comments (3)
  1. [Section IV and Discussion; Table I] Report-generation evaluation is circular. The reference reports in Table I are generated and refined by the same multi-LLM pipeline (GPT-4o-mini generation; Claude/Gemini/DeepSeek review) that produced the training targets. BLEU/ROUGE/METEOR/BERT-F1 therefore measure fit to the generator's style, not diagnostic accuracy or completeness. The Discussion's admission of a 'self-consistency bias' names the problem but does not resolve it. To support the abstract claim of better report generation, the authors should add independent evaluation: expert radiologist ratings, automatic factual checking against BraTS ground-truth labels (tumor type, modality, location), or evaluation against externally authored human reports. VQA is less affected because answers are tied to labels, but the questions and answer options are also synthetic.
  2. [Section IV-A; Tables I and II] The comparison is not controlled: LLaVA/LLaVA-Med receive a single axial slice centered on the ROI, M3D receives 32 slices, and Ours receives all axial slices. The reported gains may reflect increasing input information rather than the model or training method. Please report an ablation with the same number of slices for all models, or at least evaluate the baselines with comparable volumetric input, and state per-model input coverage explicitly. Without this, the central claim of model superiority is not established.
  3. [Tables I/II and Section IV-B, Fig. 6] The empirical evidence lacks statistical support. Tables I and II show single runs with no error bars, seeds, or significance tests despite large reported gaps. Figure 6 uses t-SNE centroid distances to claim that refinement improves separability, but t-SNE is a nonlinear embedding that does not preserve distances, and no statistical test is given. Please provide repeated runs with mean±std and appropriate significance tests, and use a distance-preserving feature-space measure in the original embedding space if refinement quality is claimed.
minor comments (6)
  1. [Fig. 1] The annotations contain OCR/encoding errors ('I nac cur ate UnnecessaryEncoding Err or', '?!!@#luding') that should be corrected.
  2. [Table I] The dimension column does not list '3D' for Ours; define all acronyms (BLEU, ROUGE, METEOR, BERT-F1) in the caption.
  3. [References] Reference [22] is given as a Semantic Scholar URL; please cite the Claude 3 model report properly.
  4. [Section II-B] The 'reviewer acceptance threshold' (2 of 3) is an arbitrary design choice; a sensitivity analysis on the number of reviewers and threshold would strengthen the refinement claim.
  5. [Section IV-B] Figure 6(b) reports centroid distances in t-SNE-reduced space; please clarify whether the distances are computed in the reduced space and also report raw embedding-space distances.
  6. [Abstract and Contributions] 'Better diagnosis and treatment' overstates what automatic text/VQA metrics can show, especially given the Discussion's acknowledged lack of clinical validation.

Circularity Check

1 steps flagged

Report-generation scores are measured against LLM-generated references from the same pipeline that produced the training data, so the headline gains partly measure fit to the generator.

specific steps
  1. fitted input called prediction [Section IV (Results, opening paragraph; Table I) and Section V (Discussion, Limitations)]
    "In this study, all evaluations, including report generation and assessment, were conducted using our previously constructed image-text paired dataset because there are no text-paired 3D MRI datasets. Further, there are currently no established benchmarks or alternative evaluation methods for directly assessing automatically generated reports on these datasets. [...] Second, both the training and evaluation data are generated and refined by LLMs, which may introduce a self-consistency bias; performance gains could partially reflect alignment within the synthetic data pipeline rather than indepe"

    The evaluation references for report generation are produced by the same multi-LLM pipeline (GPT-4o-mini generation plus Claude/Gemini/DeepSeek refinement) that produced the training targets. The VLM is fine-tuned to imitate those reports; scoring its outputs against the same generator's references with BLEU/ROUGE/METEOR/BERT-F1 therefore measures in-distribution similarity to the generator, not independent clinical quality. The paper admits this as a 'self-consistency bias,' and the abstract's central report-generation superiority claim rests on Table I, which uses these synthetic references. VQA is less affected because its labels trace to BraTS ground-truth annotations.

full rationale

The central report-generation claim is partially circular by construction: the training targets and evaluation references are both outputs of the same LLM-based generation/refinement pipeline, so high BLEU/ROUGE/METEOR/BERT-F1 values partly reflect how closely the model reproduces the generator's style and template. The Discussion's explicit concession that 'performance gains could partially reflect alignment within the synthetic data pipeline' is the key load-bearing admission. VQA evidence is more independent, since tumor type/modality/location answers are grounded in BraTS labels and masks, which keeps the overall paper from being fully circular. Self-citations in the reference list are not load-bearing for the main derivation, and no external benchmarks or expert validation are provided to break the report-generation circularity. Score 6 reflects partial circularity confined to the report-generation evaluation, which is one of the paper's two headline contributions.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new physical or architectural entities are introduced. The 'multi-LLM cooperative system' is a workflow, and the dataset is constructed from existing BraTS data with LLM-generated text, so the invented-entities ledger is empty. The main load-bearing additions are the synthetic-data assumptions listed above.

free parameters (4)
  • LoRA rank and alpha = r=32, alpha=32
    Training hyperparameters chosen by hand; no ablations are shown and the final VLM performance depends on them (Section II-C/Implementation details).
  • Two-stage epoch schedule = 5 epochs perceiver-only + 5 epochs joint fine-tuning
    Hand-set without ablations; training convergence and final performance depend on it (Implementation details).
  • Learning rate = 1e-6 (AdamW)
    Hand-set; critical for fine-tuning stability of the perceiver and LoRA weights (Implementation details).
  • Reviewer acceptance threshold = at least 2 of 3 reviewer LLMs accept
    Voting rule for the multi-LLM refinement; changes dataset composition and quality. No sensitivity analysis is provided (Section II-B).
axioms (5)
  • domain assumption LLM-generated reports and multi-LLM review produce text accurate enough to serve as training references and as evaluation ground truth.
    Core of the data pipeline; no expert validation; acknowledged in Discussion as a limitation.
  • domain assumption The three reviewing LLMs' majority vote approximates clinical expert review.
    Introduction says multi-LLM collaboration 'aims to approximate expert review'; no evidence that reviewer agreement correlates with clinical correctness.
  • domain assumption Pretrained VQ-GAN encoder (from [17]/[30]) yields 3D latent tokens that preserve clinically relevant MRI information for the LLM.
    Used in Eq. (1); no analysis of information loss or tokenization artifacts is provided.
  • domain assumption NLP overlap metrics (BLEU, ROUGE, METEOR, BERT-F1) reflect radiological report quality.
    Section IV uses them as the primary report-generation evaluation; the paper itself states this is insufficient without expert assessment.
  • domain assumption BraTS segmentation labels and brain atlas regions give a complete and accurate account of tumor location and shape for report generation.
    Data generation in Section II-A relies on these labels; any label errors would propagate to the generated text and to the trained VLM.

pith-pipeline@v1.3.0-alltime-deepseek · 9871 in / 14360 out tokens · 148849 ms · 2026-08-02T01:39:57.527829+00:00 · methodology

0 comments
read the original abstract

Recent advances in large language models (LLMs) and their extension to vision-language models (VLMs) have made it easier to combine text and images for tasks such as report generation. Existing VLMs in medicine typically focus on 2D images (chest X-rays), and their extension to 3D imaging has been difficult because of the lack of paired 3D imaging-text data. Thus, we introduce a new method for creating a 3D image-text dataset for brain oncology using 3D MRI scans of glioma and meningioma cases. We use a cooperative system in which several LLMs work together to generate and check reports, ensuring that they are accurate and clear. By leveraging the new 3D MRI-text dataset, we further build a VLM that converts MRI scans into tokens and aligns them with text instructions. Our VLM performed better in report generation and visual question answering tasks than other 2D and 3D methods. Our method not only improves the quality of reports but also helps with better diagnosis and treatment in brain oncology.

Figures

Figures reproduced from arXiv: 2607.14581 by Hyunjin Park, Jonghun Kim, Sinyoung Ra.

Figure 1
Figure 1. Figure 1: Example of Report Generation. The initial report (red box) is imprecise, includes unnecessary details, and produces hallucinations. Our Multi-LLM refinement process removes unnecessary information and captures detailed image information. The refined report (blue box) accurately generates the essential MRI details. several challenges. Unlike X-ray datasets, publicly available large-scale image-text paired d… view at source ↗
Figure 2
Figure 2. Figure 2: Instruction-driven initial report generation. Reports and VQA data are generated using basic MRI information and tumor location/shape with instructions. The tumor location is provided based on the brain atlas. The combined clinical information with instructions is fed into GPT 4o-mini to generate initial data, such as reports and VQA sets. challenges in practice due to limited expert availability, high cos… view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of training framework with a 3D image token. A 3D image is compressed into a 3D latent using a pretrained VQ encoder. Then a perceiver converts the latent into an image token, which is used to train an LLM with an instruction. The text token represents the instruction provided during training. The backbone of the LLM is frozen and only LoRA is fine-tuned. Instruction with criteria: "You are an… view at source ↗
Figure 5
Figure 5. Figure 5: Example of text generation of various VLMs. Left: Results of visual question answering. VQA is closed-ended and each model infers answers based on the input and questions. Right: Results of report generation. Each model generates a report on 3D MRI, highlighting the important and unnecessary elements in the image. TABLE I PERFORMANCE OF REPORT GENERATION. EACH MODEL IS FED AN IMAGE AND INSTRUCTIONS, AND TH… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of the text embedding for the initial and refined reports. Left: Feature space visualization with t-SNE showing clear separation by tumor type and modality. Right: Displays the distances between the centroids of the clusters identified in left. TABLE II PERFORMANCE OF VISUAL QUESTION ANSWERING. EACH MODEL RECEIVES QUESTIONS IN THREE CATEGORIES (TUMOR TYPE, MODALITY, AND TUMOR LOCATION) AND IS… view at source ↗

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

39 extracted references · 9 linked inside Pith

  1. [1]

    A survey of large language models,

    W. X. Zhaoet al., “A survey of large language models,” 2024

    2024

  2. [2]

    Advancements and applications of large language models in natural language processing: A comprehensive review,

    M. Ren, “Advancements and applications of large language models in natural language processing: A comprehensive review,”Applied and Computational Engineering, vol. 97, pp. 55–63, 2024

    2024

  3. [3]

    Enhancing radiomics features via a large language model for classifying benign and malignant breast tumors in mammography,

    S. Ra, J. Kim, I. Na, E. S. Ko, and H. Park, “Enhancing radiomics features via a large language model for classifying benign and malignant breast tumors in mammography,”Computer Methods and Programs in Biomedicine, vol. 265, p. 108765, 2025

    2025

  4. [4]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” in Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  5. [5]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” inInternational conference on machine learning. PMLR, 2022, pp. 12 888–12 900

    2022

  6. [6]

    Visual prompt tuning for task-flexible medical image synthesis,

    J. Kim and H. Park, “Visual prompt tuning for task-flexible medical image synthesis,”Computer Methods and Programs in Biomedicine, p. 109244, 2026

    2026

  7. [7]

    LLaV A-med: Training a large language-and-vision assistant for biomedicine in one day,

    C. Liet al., “LLaV A-med: Training a large language-and-vision assistant for biomedicine in one day,” inThirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023

    2023

  8. [8]

    Vision-language generative model for view-specific chest x-ray generation,

    H. Leeet al., “Vision-language generative model for view-specific chest x-ray generation,”arXiv preprint arXiv:2302.12172, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    XrayGPT: Chest radiographs summarization using large medical vision-language models,

    O. C. Thawakaret al., “XrayGPT: Chest radiographs summarization using large medical vision-language models,” inProceedings of the 23rd Workshop on Biomedical Natural Language Processing. Bangkok, Thailand: Association for Computational Linguistics, Aug. 2024, pp. 440–448

    2024

  10. [10]

    Visual instruction-finetuned language model for versatile brain mr image tasks,

    J. Kim, S. Ra, and H. Park, “Visual instruction-finetuned language model for versatile brain mr image tasks,”arXiv preprint arXiv:2604.02748, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports,

    A. E. Johnsonet al., “Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports,”Scientific data, vol. 6, no. 1, p. 317, 2019

    2019

  12. [12]

    Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison,

    J. Irvin, P. Rajpurkar, M. Ko, Y . Yu, S. Ciurea-Ilcus, C. Chute, H. Mark- lund, B. Haghgoo, R. Ball, K. Shpanskayaet al., “Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison,” inProceedings of the AAAI conference on artificial intelligence, vol. 33, no. 01, 2019, pp. 590–597

    2019

  13. [13]

    Privacy preserving chest x-ray classification in latent space with homomorphically encrypted neural inference,

    J. Kim, G. Jo, S. Ra, and H. Park, “Privacy preserving chest x-ray classification in latent space with homomorphically encrypted neural inference,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 485–495

    2025

  14. [14]

    Collaboration between clinicians and vision–language models in radiology report generation,

    R. Tannoet al., “Collaboration between clinicians and vision–language models in radiology report generation,”Nature Medicine, vol. 31, no. 2, pp. 599–608, 2025

    2025

  15. [15]

    A novel approach to eliminating hal- lucinations in large language model-assisted causal discovery,

    G. Sng, Y . Zhang, and K. Mueller, “A novel approach to eliminating hal- lucinations in large language model-assisted causal discovery,” in2024 IEEE MIT Undergraduate Research Technology Conference (URTC). IEEE, 2024, pp. 1–6

    2024

  16. [16]

    Enhancing large language models with ensemble of critics for mitigating toxicity and hallucination,

    S. Mousavi, R. L. Gutierrez, D. Rengarajan, V . Gundechaet al., “Enhancing large language models with ensemble of critics for mitigating toxicity and hallucination,” inR0-FoMo:Robustness of Few-shot and Zero-shot Learning in Large Foundation Models, 2023. [Online]. Available: https://openreview.net/forum?id=4uiOPSvbN6

    2023

  17. [17]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12 873–12 883

    2021

  18. [18]

    Biomedical visual instruction tuning with clinician prefer- ence alignment,

    H. Cuiet al., “Biomedical visual instruction tuning with clinician prefer- ence alignment,” inThe Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024

    2024

  19. [19]

    An anatomically comprehensive atlas of the adult human brain transcriptome,

    M. J. Hawrylyczet al., “An anatomically comprehensive atlas of the adult human brain transcriptome,”Nature, vol. 489, no. 7416, pp. 391– 399, 2012

    2012

  20. [20]

    An automated labeling system for subdividing the human cerebral cortex on mri scans into gyral based regions of interest,

    R. S. Desikanet al., “An automated labeling system for subdividing the human cerebral cortex on mri scans into gyral based regions of interest,” Neuroimage, vol. 31, no. 3, pp. 968–980, 2006

    2006

  21. [21]

    Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach,

    H. J. Aertset al., “Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach,”Nature communications, vol. 5, no. 1, p. 4006, 2014

    2014

  22. [22]

    The claude 3 model family: Opus, sonnet, haiku

    “The claude 3 model family: Opus, sonnet, haiku.” [Online]. Available: https://api.semanticscholar.org/CorpusID:268232499

  23. [23]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context,

    G. Teamet al., “Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context,”arXiv preprint arXiv:2403.05530, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Deepseek llm: Scaling open-source language models with longtermism,

    X. Biet al., “Deepseek llm: Scaling open-source language models with longtermism,”arXiv preprint arXiv:2401.02954, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Gpt-4 technical report,

    J. Achiamet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification,

    U. Baidet al., “The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification,”arXiv preprint arXiv:2107.02314, 2021

    Pith/arXiv arXiv 2021

  27. [27]

    The asnr-miccai brain tumor segmentation (brats) challenge 2023: Intracranial meningioma,

    D. LaBellaet al., “The asnr-miccai brain tumor segmentation (brats) challenge 2023: Intracranial meningioma,” 2023

    2023

  28. [28]

    M3d: Advancing 3d medical image analysis with multi-modal large language models,

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

    Pith/arXiv arXiv 2024

  29. [29]

    E3d-gpt: Enhanced 3d visual foundation for medical vision-language model,

    H. Laiet al., “E3d-gpt: Enhanced 3d visual foundation for medical vision-language model,”arXiv preprint arXiv:2410.14200, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Adaptive latent diffusion model for 3d medical image to image translation: Multi-modal magnetic resonance imaging study,

    J. Kim and H. Park, “Adaptive latent diffusion model for 3d medical image to image translation: Multi-modal magnetic resonance imaging study,” inProceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, 2024, pp. 7604–7613

    2024

  31. [31]

    Lora: Low-rank adaptation of large language models,

    E. J. Huet al., “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685, 2021

    Pith/arXiv arXiv 2021

  32. [32]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,

    W.-L. Chianget al., “Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,”See https://vicuna. lmsys. org (accessed 14 April 2023), vol. 2, no. 3, p. 6, 2023

    2023

  33. [33]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inInternational Conference on Learning Representations, 2019

    2019

  34. [34]

    Bleu: a method for automatic evaluation of machine translation,

    K. Papineniet al., “Bleu: a method for automatic evaluation of machine translation,” inProceedings of the 40th annual meeting of the Associa- tion for Computational Linguistics, 2002, pp. 311–318

    2002

  35. [35]

    Rouge: A package for automatic evaluation of summaries,

    C.-Y . Lin, “Rouge: A package for automatic evaluation of summaries,” inText summarization branches out, 2004, pp. 74–81

    2004

  36. [36]

    Meteor: An automatic metric for mt evalua- tion with improved correlation with human judgments,

    S. Banerjee and A. Lavie, “Meteor: An automatic metric for mt evalua- tion with improved correlation with human judgments,” inProceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 2005, pp. 65–72

    2005

  37. [37]

    Bertscore: Evaluating text generation with bert,

    T. Zhang*et al., “Bertscore: Evaluating text generation with bert,” in International Conference on Learning Representations, 2020

    2020

  38. [38]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11 2019

    2019

  39. [39]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.”Journal of machine learning research, vol. 9, no. 11, 2008

    2008