pith. sign in

arxiv: 2607.01814 · v1 · pith:NJ6IHRQYnew · submitted 2026-07-02 · 💻 cs.AI

MMIR-TCM: Memory-Integrated Multimodal Inference and Retrieval for TCM Clinical Decision Support

Lihui Luo , Joongwon Chae , Ziyan Chen , Yang Liu , Siyi Cheng , Weihan Gao , Zelin Zeng , Xiaoming Yin
show 8 more authors
Samaneh Beheshti Kashi Dongmei Yu Lian Zhang Jing Sui Zeming Liang Jiansong Ji Peter E. Lobie Peiwu Qin
This is my paper

Pith reviewed 2026-07-03 13:37 UTC · model grok-4.3

classification 💻 cs.AI
keywords Traditional Chinese Medicinemultimodal large language modelsretrieval-augmented generationtongue diagnosisclinical decision supportMedTCM datasetTDEU metricmemory-augmented segmentation
0
0 comments X

The pith

MMIR-TCM integrates memory-augmented segmentation, fine-tuned multimodal inference, and retrieval-augmented generation to outperform GPT-4o and Gemini 2.5 Flash on TCM tasks.

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

The paper aims to reduce subjectivity and the visual-textual gap in TCM tongue diagnosis and syndrome differentiation by building a system that copies expert workflows. MMIR-TCM processes images through memory-enhanced extraction, produces structured diagnoses, and grounds prescriptions in retrieved evidence, all on a newly created MedTCM dataset. A custom TDEU metric is introduced because standard scores miss clinical importance. Experiments show the framework delivers higher performance than current top models on these tasks.

Core claim

MMIR-TCM emulates the diagnostic process of TCM experts through a three-stage architecture that combines a training-free Memory-SAM module for robust tongue extraction, a fine-tuned Qwen3-VL model for structured tongue diagnosis generation, and a Qwen3-based RAG component for evidence-grounded clinical decision support, all validated on the MedTCM dataset with the TDEU metric.

What carries the argument

The three-stage MMIR-TCM architecture that uses memory-augmented segmentation together with multimodal generation and retrieval-augmented generation to connect visual tongue features to textual clinical reasoning.

If this is right

  • TCM syndrome differentiation from tongue images becomes more reproducible across different practitioners.
  • Prescription generation gains reliability through direct evidence retrieval rather than model recall alone.
  • The MedTCM dataset can serve as a shared benchmark for testing future multimodal TCM systems.
  • TDEU provides a practical alternative for scoring medical outputs when generic metrics fall short.
  • Clinical decision support tools for TCM can be assembled without full retraining of large base models.

Where Pith is reading between the lines

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

  • The memory component could support consistent tracking of patient changes over multiple visits.
  • Similar memory-plus-retrieval patterns might transfer to other image-to-text medical domains such as skin or eye imaging.
  • The outputs could supply standardized examples for TCM education or remote consultation support.
  • Linking the RAG store to patient history records could strengthen long-term decision quality.

Load-bearing premise

That the TDEU metric validly measures clinical accuracy and diagnostic importance where standard metrics fail, and that the MedTCM dataset provides a representative benchmark for real-world TCM tasks.

What would settle it

A panel of TCM experts rating outputs from MMIR-TCM no higher than those from GPT-4o on diagnostic accuracy and importance for a fresh set of tongue images would falsify the performance claim.

read the original abstract

Traditional Chinese Medicine (TCM) diagnosis, particularly through tongue inspection, faces persistent challenges in subjectivity and reproducibility. The application of multimodal artificial intelligence to TCM clinical tasks, such as syndrome differentiation and prescription generation, is significantly hampered by the semantic gap between visual tongue features and textual reasoning, as well as the lack of large-scale, standardized datasets. To address these challenges, we introduce MMIR-TCM, a novel framework that emulates the diagnostic process of TCM experts by integrating multimodal large language model(MLLM) with memory-augmented segmentation and retrieval-augmented generation (RAG). Employing a three-stage architecture, MMIR-TCM integrates a training-free Memory-SAM module for robust tongue extraction, a fine-tuned Qwen3-VL model for structured tongue diagnosis generation, and a Qwen3-based RAG component for evidence-grounded clinical decision support generation. The framework was developed and validated using MedTCM, a new large-scale multimodal dataset that we introduce specifically for advanced TCM research. To properly evaluate our framework's clinical accuracy, which existing metrics fail to capture, we also developed TDEU, a domain-specific evaluation metric incorporating semantic understanding and diagnostic importance. Our comprehensive experiments demonstrate that MMIR-TCM significantly outperforms leading models, including GPT-4o and Gemini 2.5 Flash.

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

3 major / 1 minor

Summary. The manuscript introduces MMIR-TCM, a three-stage multimodal framework for TCM clinical decision support that integrates a training-free Memory-SAM module for tongue extraction, a fine-tuned Qwen3-VL model for structured tongue diagnosis generation, and a Qwen3-based RAG component for evidence-grounded prescription generation. It presents the new MedTCM multimodal dataset and the custom TDEU evaluation metric, claiming that comprehensive experiments show MMIR-TCM significantly outperforms leading models including GPT-4o and Gemini 2.5 Flash.

Significance. If the outperformance claims are substantiated with rigorous quantitative evidence and the TDEU metric is validated as correlating with clinical accuracy, the work could meaningfully advance multimodal AI applications in Traditional Chinese Medicine by addressing the visual-textual semantic gap. The introduction of a domain-specific dataset and metric tailored to TCM tasks represents a potentially useful contribution, but the current absence of supporting results, definitions, and validation data prevents determination of its actual significance.

major comments (3)
  1. [Abstract] Abstract: the central claim that 'MMIR-TCM significantly outperforms leading models, including GPT-4o and Gemini 2.5 Flash' is asserted without any quantitative results, tables, baselines, error analysis, or statistical comparisons. The central contribution therefore rests on an unevidenced assertion.
  2. [Abstract] Abstract: TDEU is introduced as a domain-specific metric that incorporates semantic understanding and diagnostic importance where existing metrics fail, yet no definition, formula, weighting scheme, computation details, or validation (e.g., expert correlation or inter-annotator agreement) is supplied.
  3. [Abstract] Abstract: MedTCM is presented as a new large-scale multimodal dataset developed specifically for advanced TCM research, but the manuscript provides no information on dataset size, collection protocol, annotation process, representativeness checks, or any clinical validation steps.
minor comments (1)
  1. The abstract refers to 'comprehensive experiments' and 'our framework was developed and validated using MedTCM' but the manuscript does not appear to contain the corresponding results, ablation studies, or dataset statistics sections that would normally support these statements.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We agree that the abstract should better substantiate its claims with concrete details and will revise it to include key quantitative results, a brief definition of TDEU, and essential dataset information. The full experimental results, metric definitions, and dataset documentation are provided in the main body of the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'MMIR-TCM significantly outperforms leading models, including GPT-4o and Gemini 2.5 Flash' is asserted without any quantitative results, tables, baselines, error analysis, or statistical comparisons. The central contribution therefore rests on an unevidenced assertion.

    Authors: We agree that the abstract would be strengthened by including specific quantitative evidence. In the revised version, we will add key performance numbers (e.g., TDEU scores for MMIR-TCM versus GPT-4o and Gemini 2.5 Flash) and reference the corresponding tables and statistical comparisons that appear in the experimental section. revision: yes

  2. Referee: [Abstract] Abstract: TDEU is introduced as a domain-specific metric that incorporates semantic understanding and diagnostic importance where existing metrics fail, yet no definition, formula, weighting scheme, computation details, or validation (e.g., expert correlation or inter-annotator agreement) is supplied.

    Authors: We acknowledge the need for a concise definition in the abstract. We will revise the abstract to include a short description of TDEU's formulation and its emphasis on diagnostic importance. The complete formula, weighting scheme, computation procedure, and validation results (including expert correlation) are already detailed in the dedicated evaluation-metric section of the manuscript. revision: yes

  3. Referee: [Abstract] Abstract: MedTCM is presented as a new large-scale multimodal dataset developed specifically for advanced TCM research, but the manuscript provides no information on dataset size, collection protocol, annotation process, representativeness checks, or any clinical validation steps.

    Authors: We agree that the abstract should briefly convey the scale and construction of MedTCM. We will add summary statistics on dataset size, collection and annotation protocols, and validation steps to the abstract. Full details on these aspects, including representativeness checks and clinical validation, are provided in the dataset section of the paper. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical claims rest on introduced artifacts without definitional reduction

full rationale

The paper presents an empirical framework (MMIR-TCM) evaluated on a newly introduced dataset (MedTCM) and metric (TDEU). The abstract describes TDEU as a domain-specific metric developed because existing metrics fail to capture clinical accuracy, but supplies no equations, fitting procedure, or definition that reduces TDEU scores to the model's architecture or outputs by construction. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing premises. The outperformance claim is therefore an empirical statement on self-created benchmarks rather than a derivation that collapses to its inputs; this is the normal case for applied ML papers introducing new resources and does not meet the criteria for any enumerated circularity pattern.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no visible free parameters, axioms, or invented entities; the central claim rests on the unstated assumption that the custom TDEU metric and MedTCM dataset are appropriate and unbiased.

pith-pipeline@v0.9.1-grok · 5832 in / 1098 out tokens · 51674 ms · 2026-07-03T13:37:33.227185+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

82 extracted references · 82 canonical work pages · 5 internal anchors

  1. [1]

    We propose MMIR-TCM, a memory-integrated mul- timodal pipeline for TCM clinical decision sup- port that integrates training free tongue extrac- tion, attribute-level tongue diagnosis generation, and retrieval-augmented prescription generation to produce evidence-grounded clinical decision support

    work page

  2. [2]

    The dataset will be made publicly available and maintained

    We construct MedTCM, a large scale multimodal TCM dataset with tongue images, diagnostic re- ports, and clinical prescriptions collected from mul- tiple hospitals to ensure diversity and clinical rele- vance. The dataset will be made publicly available and maintained

    work page

  3. [3]

    We develop TDEU, a domain aware evaluation met- ric for tongue diagnosis that incorporates semantic similarity and clinical importance, addressing the limitations of traditional text matching metrics. II. R ELATED WORK A. TCM-Specific Large Language Models Recent efforts have adapted large language models (LLMs) to TCM through domain-specific pretraining ...

    work page

  4. [4]

    Tongue Extractor. A training-free Memory-SAM segmenter retrieves visual exemplars and converts them into structured foreground/background point prompts for SAM2 [42], yielding a precise tongue mask and a background-removed region-of-interest (ROI) I ROI tongue

    work page

  5. [5]

    Tongue Diagnosis Generator. A fine-tuned Qwen3- VL model analyzes I ROI tongue and outputs a one- sentence, attribute-level report D that consistently enumerates tongue body color/shape and coating color/thickness/texture with anatomical locations

    work page

  6. [6]

    Prescription Generator. A Qwen3-based generator fuses D with Mpatient, retrieves similar cases from a de-identified EMR knowledge base via vector search, and synthesizes syndrome differentiation and an herbal prescription with concise evidence-based reasoning. This modular design enables independent optimization of each stage while preserving end-to-end t...

    work page

  7. [7]

    Diagnosis:

    The above four lines must be output, each starting with "Diagnosis:", "Dialectics:", "Prescription:", "Reason for Diagnosis:“

    work page

  8. [8]

    The diagnosis is concise and clear, consisting of 2 -8 words

    work page

  9. [9]

    Dialectically accurate, 4-8 words

    work page

  10. [10]

    Separate each medication in the prescription with "" (three spaces) to avoid duplicate medication

    work page

  11. [11]

    Must include the section on usage and dosage

    work page

  12. [12]

    The diagnostic reasons should be detailed, including symptom analysis, tongue and pulse analysis, and pathogenesis explanation

    work page

  13. [13]

    2 : The detailed prompt template used in MMIR- TCM

    Do not output thought processes, explanations, or any other content COT Prompt for MMIR- TCM (a) (b) Fig. 2 : The detailed prompt template used in MMIR- TCM. a) Dense Feature Extraction. : Given a tongue image Itongue, we extract patch embeddings using a pretrained DINOv3 (ViT-L/16) encoder [44]. Each patch embedding is ℓ2-normalized to form a dense featu...

    work page

  14. [14]

    Model and Training Objective : We employ Qwen3-VL- 30B [46], a vision-language model with a vision encoder (SigLIP), projection layer, and autoregressive language decoder. Given a tongue image Itongue and target report T , we minimize the cross-entropy loss: L(θ) = − ∑ (I,T )∈D log Pθ(T | I), (5) where D is our training dataset of 2,805 image-report pairs...

    work page

  15. [15]

    Structured Output Format : We enforce a standardized output schema through system prompts (See Fig. 2 (a)). Example output: ”Tongue body pale red, shape normal, coating white, thin, slightly greasy, distributed evenly with mild redness at tip. ” This format ensures that every report covers the same diagnostic dimensions, enabling reliable parsing and stru...

    work page

  16. [16]

    Input images are resized to 768× 768 pixels

    Implementation Details : We use Qwen3-VL-30B with LoRA applied to both vision and language components (rank 64, 10 epochs, dropout = 0); the optimizer uses a learning rate of 5×10−5. Input images are resized to 768× 768 pixels. During inference, we use nucleus sampling with p = 0 .9 and temperature T = 0 .7 to balance consistency and natural language flue...

    work page

  17. [17]

    The case bank is built from 124,593 anonymized patient records collected from four major TCM hospitals

    Knowledge Base Construction and Indexing : The foun- dation of our RAG system is a comprehensive Memory Base, consisting of two key components: the LiuJing Theory and the Clinical Case Bank. The case bank is built from 124,593 anonymized patient records collected from four major TCM hospitals. Each record contains complete clinical information—including c...

    work page

  18. [18]

    Document Processing: Each record in memory base is structured as a document containing multiple fields (chief complaint, history, pulse, tongue, syn- drome, prescription)

    work page

  19. [19]

    Semantic Embedding: Documents are encoded into 768-dimensional dense vectors using a domain- adapted Chinese medical language model

    work page

  20. [20]

    This indexing strategy ensures both retrieval accuracy and computational efficiency, critical for real-time clinical decision support applications

    Hierarchical Indexing: F AISS constructs hierarchi- cal navigable small world (HNSW) graphs [48] for approximate nearest neighbor search, enabling sub- linear query time complexity. This indexing strategy ensures both retrieval accuracy and computational efficiency, critical for real-time clinical decision support applications

    work page

  21. [21]

    The retrieval module then identifies the k most clinically relevant cases using cosine similarity: R = Retrieve(q, k) = {C1, C2,

    Query Construction and Retrieval : Given the tongue diagnosis output Ttongue from the visual analysis module and patient metadata Mpatient = ( Q, H, P ), we construct a unified query vector: q = Embed(Ttongue ⊕ Mpatient) (6) where ⊕ denotes information concatenation, and Embed (·) represents the semantic embedding function mapping text to a dense vector s...

    work page

  22. [22]

    Following the prompt shown in Fig

    Context-Aware Diagnostic Generation : The final stage of the prescription generator synthesizes the retrieved clinical cases with the current patient’s information to generate comprehensive diagnostic outputs. Following the prompt shown in Fig. 2 (b), the Qwen3-based reasoning engine processes an evidence-augmented context formed by combining the retrieve...

    work page

  23. [23]

    Token Decomposition and Categorization : Given a pre- dicted tongue report ˆT and reference report T ∗, we parse both into atomic attribute tokens using predefined pattern-matching rules. Each token is assigned to one of four categories: • TONGUE: tongue body color and shape • COAT: coating color, thickness, and texture • LOCATION: anatomical regions • OT...

    work page

  24. [24]

    Each token follows the format PREFIX_V ALUE (e.g., COLOR_red, THICK_thin)

    Semantic Similarity and Optimal Matching : For each category c, we compute pairwise semantic similarities be- tween tokens in Pc and Lc. Each token follows the format PREFIX_V ALUE (e.g., COLOR_red, THICK_thin). Similarity S(p, l) is defined through a hierarchical match- ing scheme: S(p, l) =    1.0 if exact match Synπ(p)(ν(p), ν(l)) if π(p) = π(...

    work page

  25. [25]

    Through consultation with senior TCM practitioners, we let wTONGUE=2.0, wCOAT=1.5, wLOCATION=1.0, and wOTHER=0.5

    Clinical Importance Weighting : We evaluate four cat- egories: TONGUE, COAT, LOCATION, and OTHER (auxiliary descriptors). Through consultation with senior TCM practitioners, we let wTONGUE=2.0, wCOAT=1.5, wLOCATION=1.0, and wOTHER=0.5. Define the category set C = {TONGUE, COAT, LOCATION, OTHER}. The overall score is a weighted average: TDEU = ∑ c∈C wc · T...

    work page 2025

  26. [26]

    4, was structured around a primary institution leading data standardization, complemented by three partner hospitals contributing di- verse clinical cases

    Multi-Center Data Collection Strategy : The data col- lection framework, illustrated in Fig. 4, was structured around a primary institution leading data standardization, complemented by three partner hospitals contributing di- verse clinical cases. This multi-center strategy was crucial for ensuring data diversity and mitigating biases inherent in single-...

    work page 2021

  27. [27]

    few-annotation

    Dataset Statistics and Characteristics : As shown in Table I, the dataset exhibits rich demographic diver- sity and comprehensive syndrome coverage. The tongue image subset (2,805 pairs) was selected to ensure high image quality. Expert TCM physicians verified all image- diagnosis pairs for consistency and clinical accuracy. The dataset will be continuous...

    work page

  28. [28]

    Ablation on Tongue Extractor : In deep learning- based tongue diagnosis and medical imaging, it is well- established that mask segmentation (background removal) or ROI standardization enhances model performance and stability. This is generally attributed to two factors: (i) suppressing non-medical sources of variance such as background clutter, illuminati...

    work page

  29. [29]

    LoRA adapters are inserted into the linear layers of both the self-attention modules ( q, k, v, o) and the feed-forward modules (gate, up, down)

    Ablation on Fine-Tuning Strategy and Hyperparameters : The model architecture is composed of three main compo- nents: (i) a vision tower, (ii) a multimodal projector, and (iii) a LLaMA-style language transformer. LoRA adapters are inserted into the linear layers of both the self-attention modules ( q, k, v, o) and the feed-forward modules (gate, up, down)...

    work page

  30. [30]

    We systematically re- moved specific modules from the full model and evaluated the performance degradation

    Ablation on Prescription Generator : To validate the effectiveness and contribution of each key component of our proposed MMIR-TCM’s RAG framework, we con- ducted a series of ablation studies. We systematically re- moved specific modules from the full model and evaluated the performance degradation. The results, visualized as heatmaps in Fig. 7, compare t...

    work page

  31. [31]

    Outputs were scored from 1 to 5 on five dimensions: Acc_SN, Acc_Diag, Validity, Safety, and Rationality

    Overall Ratings : We conducted a blinded preference study with 12 experienced TCM clinicians comparing MMIR-TCM and GPT-4o on 20 random cases (240 evaluations). Outputs were scored from 1 to 5 on five dimensions: Acc_SN, Acc_Diag, Validity, Safety, and Rationality. Aggregated ratings (Fig. 8) show a consis- tent clinician preference for MMIR-TCM, with fee...

    work page

  32. [32]

    hallucinations

    Challenges and Limitations : To quantify real-world accuracy, we performed manual error analysis against formal TCM hospital records. Five experts reviewed 2,362 generated cases and re-evaluated a random 50% subset (1,181) for disease diagnosis, meridian attribution, and syndrome consistency. Inconsistencies were marked as incomplete or incorrect, and 50 ...

    work page

  33. [33]

    Vision-language foundation models for medical imaging: a review of current practices and innovations,

    J. S. Ryu, H. Kang, Y. Chu, and S. Yang, “Vision-language foundation models for medical imaging: a review of current practices and innovations,” Biomedical Engineering Letters, pp. 1–22, 2025

    work page 2025

  34. [34]

    Tongue diagnosis indices for gas- troesophageal reflux disease: a cross-sectional, case-controlled observational study,

    T.-C. Wu, C.-N. Lu, W.-L. Hu, K.-L. Wu, J. Y. Chiang, J.- M. Sheen, and Y.-C. Hung, “Tongue diagnosis indices for gas- troesophageal reflux disease: a cross-sectional, case-controlled observational study,” Medicine, vol. 99, no. 29, p. e20471, 2020

    work page 2020

  35. [35]

    Objective evaluation of tongue diagnosis ability using a tongue diagnosis e-learning/e-assessment system based on a standardized tongue image database,

    M. Segawa, N. Iizuka, H. Ogihara, K. Tanaka, H. Nakae, K. Usuku, K. Yamaguchi, K. Wada, A. Uchizono, Y. Nakamura et al., “Objective evaluation of tongue diagnosis ability using a tongue diagnosis e-learning/e-assessment system based on a standardized tongue image database,” Frontiers in Medical Technology, vol. 5, p. 1050909, 2023

    work page 2023

  36. [36]

    A systematic literature review and classification of knowledge discovery in traditional medicine,

    G. Arji, R. Safdari, H. Rezaeizadeh, A. Abbassian, M. Mokhtaran, and M. H. Ayati, “A systematic literature review and classification of knowledge discovery in traditional medicine,” Computer methods and programs in biomedicine, vol. 168, pp. 39–57, 2019

    work page 2019

  37. [37]

    A survey of artificial intelligence in tongue image for disease diagnosis and syndrome differentiation,

    Q. Liu, Y. Li, P. Yang, Q. Liu, C. Wang, K. Chen, and Z. Wu, “A survey of artificial intelligence in tongue image for disease diagnosis and syndrome differentiation,” Digital Health, vol. 9, p. 20552076231191044, 2023

    work page 2023

  38. [38]

    Pregenerator: Tcm prescription recommendation model based on retrieval and generation method,

    Z. Zhao, X. Ren, K. Song, Y. Qiang, J. Zhao, J. Zhang, and P. Han, “Pregenerator: Tcm prescription recommendation model based on retrieval and generation method,” IEEE Access, vol. 11, pp. 103 679–103 692, 2023

    work page 2023

  39. [39]

    Large language models in traditional chinese medicine: A short survey and outlook,

    R. Zhang, D. Tian, and Y. Wang, “Large language models in traditional chinese medicine: A short survey and outlook,” AI Medicine, vol. 2, no. 1, p. 3, 2025. 14 IEEE TRANSACTIONS AND JOURNALS TEMPLATE 1.00 | 24 0.30 | 9 0.07 | 2 0.00 | 0 0.35 | 16 0.04 | 1 0.30 | 9 1.00 | 15 0.25 | 4 0.17 | 6 0.15 | 7 0.00 | 0 0.07 | 2 0.25 | 4 1.00 | 5 0.11 | 3 0.00 | 0 0...

    work page 2025

  40. [40]

    Clinicalrag: enhancing clinical decision support through heterogeneous knowledge retrieval,

    Y. Lu, X. Zhao, and J. Wang, “Clinicalrag: enhancing clinical decision support through heterogeneous knowledge retrieval,” in Proceedings of the 1st Workshop on Towards Knowledgeable Language Models (KnowLLM 2024), 2024, pp. 64–68

    work page 2024

  41. [41]

    Tcmchat: A generative large language model for traditional chinese medicine,

    Y. Dai, X. Shao, J. Zhang, Y. Chen, Q. Chen, J. Liao, F. Chi, J. Zhang, and X. Fan, “Tcmchat: A generative large language model for traditional chinese medicine,” Pharmacological Re- search, vol. 210, p. 107530, 2024

    work page 2024

  42. [42]

    Tcmbench: A comprehensive benchmark for evaluating large language models in traditional chinese medicine,

    W. Yue, X. Wang, W. Zhu, M. Guan, H. Zheng, P. Wang, C. Sun, and X. Ma, “Tcmbench: A comprehensive benchmark for evaluating large language models in traditional chinese medicine,” arXiv preprint arXiv:2406.01126, 2024

    work page arXiv 2024

  43. [43]

    Tcmd: A traditional chinese medicine qa dataset for evaluating large language models,

    P. Yu, K. Song, F. He, M. Chen, and J. Lu, “Tcmd: A traditional chinese medicine qa dataset for evaluating large language models,” arXiv preprint arXiv:2406.04941, 2024

    work page arXiv 2024

  44. [44]

    Biancang: a tradi- tional chinese medicine large language model,

    S. Wei, X. Peng, Y. Wang, T. Shen, J. Si, W. Zhang, F. Zhu, A. V. Vasilakos, W. Lu, X. Wu et al., “Biancang: a tradi- tional chinese medicine large language model,” IEEE Journal of Biomedical and Health Informatics, 2025

    work page 2025

  45. [45]

    Tcm-ftp: fine-tuning large language models for herbal prescription prediction,

    X. Zhou, X. Dong, C. Li, Y. Bai, Y. Xu, K. C. Cheung, S. Se, X. Song, R. Zhang, X. Zho et al., “Tcm-ftp: fine-tuning large language models for herbal prescription prediction,” in 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2024, pp. 4092–4097

    work page 2024

  46. [46]

    Tcm-kllama: Intelligent generation model for traditional chinese medicine prescriptions based on knowledge graph and large language model,

    Y. Zhuang, L. Yu, N. Jiang, and Y. Ge, “Tcm-kllama: Intelligent generation model for traditional chinese medicine prescriptions based on knowledge graph and large language model,” Comput- ers in Biology and Medicine, vol. 189, p. 109887, 2025

    work page 2025

  47. [47]

    Tonguenet: accurate localization and segmentation for tongue images using deep neural net- works,

    C. Zhou, H. Fan, and Z. Li, “Tonguenet: accurate localization and segmentation for tongue images using deep neural net- works,” IEEE access, vol. 7, pp. 148 779–148 789, 2019

    work page 2019

  48. [48]

    Cross-modal attention model integrating tongue images and descriptions: a novel intelligent tcm approach for pathological organ diagnosis,

    Q. Gan, C. Wang, Z. Zhong, J. Wu, Q. Ge, L. Shi, J. Shang, and C. Liu, “Cross-modal attention model integrating tongue images and descriptions: a novel intelligent tcm approach for pathological organ diagnosis,” Frontiers in Physiology, vol. 16, p. 1580985, 2025

    work page 2025

  49. [49]

    Tcm-tongue: A standardized tongue image dataset with pathological annotations for ai- assisted tcm diagnosis,

    X. Jin, L. Gao, A. Tong, Z. Chen, J. Kong, N. Sun, H. Ma, Q. Wang, Y. Bai, and T. Su, “Tcm-tongue: A standardized tongue image dataset with pathological annotations for ai- assisted tcm diagnosis,” arXiv preprint arXiv:2507.18288, 2025

    work page arXiv 2025

  50. [50]

    Tcmeval-sdt: a benchmark dataset for syndrome differentiation thought of traditional chinese medicine,

    Z. Wang, M. Hao, S. Peng, Y. Huang, Y. Lu, K. Yao, X. Yang, and Y. Zhu, “Tcmeval-sdt: a benchmark dataset for syndrome differentiation thought of traditional chinese medicine,” Scien- tific Data, vol. 12, no. 1, p. 437, 2025

    work page 2025

  51. [51]

    Automatic tongue image quality assessment using a multi-task deep learning model,

    H. Xian, Y. Xie, Z. Yang, L. Zhang, S. Li, H. Shang, W. Zhou, and H. Zhang, “Automatic tongue image quality assessment using a multi-task deep learning model,” Frontiers in Physiology, vol. 13, p. 966214, 2022

    work page 2022

  52. [52]

    Tongue crack recognition using segmentation based deep learning,

    J. Yan, J. Cai, Z. Xu, R. Guo, W. Zhou, H. Yan, Z. Xu, and Y. Wang, “Tongue crack recognition using segmentation based deep learning,” Scientific Reports, vol. 13, no. 1, p. 511, 2023

    work page 2023

  53. [53]

    A novel tongue coating segmentation method based on improved transunet,

    J. Wu, Z. Li, Y. Cai, H. Liang, L. Zhou, M. Chen, and J. Guan, “A novel tongue coating segmentation method based on improved transunet,” Sensors, vol. 24, no. 14, p. 4455, 2024

    work page 2024

  54. [54]

    Tonguesam: An universal tongue segmentation model based on sam with zero-shot,

    S. Cao, Q. Wu, and L. Ma, “Tonguesam: An universal tongue segmentation model based on sam with zero-shot,” in 2023 IEEE international conference on bioinformatics and biomedicine (BIBM). IEEE, 2023, pp. 4520–4526

    work page 2023

  55. [55]

    Retrieval-Augmented Generation for Large Language Models: A Survey

    Y. Gao, Y. Xiong, X. Gao, K. Jia, J. Pan, Y. Bi, Y. Dai, J. Sun, H. Wang, and H. Wang, “Retrieval-augmented gen- eration for large language models: A survey,” arXiv preprint arXiv:2312.10997, vol. 2, no. 1, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  56. [56]

    Lins: A general medical q&a framework for enhancing the quality and credibility of llm- generated responses,

    S. Wang, F. Zhao, D. Bu, Y. Lu, M. Gong, H. Liu, Z. Yang, X. Zeng, Z. Yuan, B. Wan et al., “Lins: A general medical q&a framework for enhancing the quality and credibility of llm- generated responses,” Nature Communications, vol. 16, no. 1, p. 9076, 2025

    work page 2025

  57. [57]

    Opentcm: a graphrag-empowered llm-based system for traditional chinese medicine knowledge retrieval and diagnosis,

    J. He, Y. Guo, L. K. Lam, W. Leung, L. He, Y. Jiang, C. C. Wang, G. Xing, and H. Chen, “Opentcm: a graphrag-empowered llm-based system for traditional chinese medicine knowledge retrieval and diagnosis,” arXiv preprint arXiv:2504.20118, 2025

    work page arXiv 2025

  58. [58]

    Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis,

    F. Gaber, M. Shaik, F. Allega, A. J. Bilecz, F. Busch, K. Goon, V. Franke, and A. Akalin, “Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis,” npj Digital Medicine, vol. 8, no. 1, p. 263, 2025

    work page 2025

  59. [59]

    P. U. Unschuld and H. Tessenow, Huang Di Nei Jing Su Wen: AUTHOR et al.: TITLE 15 An annotated translation of Huang Di’s inner classic–basic questions: 2 volumes. Univ of California Press, 2011

    work page 2011

  60. [60]

    Enhancing large language models for clinical decision support by incorporating clinical practice guidelines,

    D. Oniani, X. Wu, S. Visweswaran, S. Kapoor, S. Kooragayalu, K. Polanska, and Y. Wang, “Enhancing large language models for clinical decision support by incorporating clinical practice guidelines,” in 2024 IEEE 12th International Conference on Healthcare Informatics (ICHI). IEEE, 2024, pp. 694–702

    work page 2024

  61. [61]

    Mtcmb: A multi-task bench- mark framework for evaluating llms on knowledge, reasoning, and safety in traditional chinese medicine,

    S. Kong, X. Yang, Y. Wei, Z. Wang, H. Tang, J. Qin, S. Lan, Y. Wang, J. Bai, Z. Chen et al., “Mtcmb: A multi-task bench- mark framework for evaluating llms on knowledge, reasoning, and safety in traditional chinese medicine,” arXiv preprint arXiv:2506.01252, 2025

    work page arXiv 2025

  62. [62]

    Tcm-ladder: A benchmark for multimodal question answering on traditional chinese medicine,

    J. Xie, Y. Yu, Z. Zhang, S. Zeng, J. He, A. Vasireddy, X. Tang, C. Guo, L. Zhao, C. Jing et al., “Tcm-ladder: A benchmark for multimodal question answering on traditional chinese medicine,” arXiv preprint arXiv:2505.24063, 2025

    work page arXiv 2025

  63. [63]

    Ragas: Automated evaluation of retrieval augmented generation,

    S. Es, J. James, L. E. Anke, and S. Schockaert, “Ragas: Automated evaluation of retrieval augmented generation,” in Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: System Demon- strations, 2024, pp. 150–158

    work page 2024

  64. [64]

    Grouse: A benchmark to evaluate evaluators in grounded question answering,

    S. Muller, A. Loison, B. Omrani, and G. Viaud, “Grouse: A benchmark to evaluate evaluators in grounded question answering,” arXiv preprint arXiv:2409.06595, 2024

    work page arXiv 2024

  65. [65]

    Medcite: Can language models generate verifiable text for medicine?

    X. Wang, M. Tan, Q. Jin, G. Xiong, Y. Hu, A. Zhang, Z. Lu, and M. Zhang, “Medcite: Can language models generate verifiable text for medicine?” arXiv preprint arXiv:2506.06605, 2025

    work page arXiv 2025

  66. [66]

    Out-of-distribution detection in medical image analysis: A survey.arXiv preprint arXiv:2404.18279, 2024

    Z. Hong, Y. Yue, Y. Chen, L. Cong, H. Lin, Y. Luo, M. H. Wang, W. Wang, J. Xu, X. Yang et al., “Out-of-distribution detection in medical image analysis: A survey,” arXiv preprint arXiv:2404.18279, 2024

    work page arXiv 2024

  67. [67]

    A scoping review of robustness con- cepts for machine learning in healthcare,

    A. Balendran, C. Beji, F. Bouvier, O. Khalifa, T. Evgeniou, P. Ravaud, and R. Porcher, “A scoping review of robustness con- cepts for machine learning in healthcare,” npj Digital Medicine, vol. 8, no. 1, p. 38, 2025

    work page 2025

  68. [68]

    Trustworthy clinical ai solutions: A unified review of uncer- tainty quantification in deep learning models for medical image analysis

    B. Lambert, F. Forbes, S. Doyle, H. Dehaene, and M. Dojat, “Trustworthy clinical ai solutions: A unified review of uncer- tainty quantification in deep learning models for medical image analysis. ” Artif. Intell. Medicine, vol. 150, p. 102830, 2024

    work page 2024

  69. [69]

    Empirical validation of conformal prediction for trustworthy skin lesions classification,

    J. Fayyad, S. Alijani, and H. Najjaran, “Empirical validation of conformal prediction for trustworthy skin lesions classification,” Computer Methods and Programs in Biomedicine, vol. 253, p. 108231, 2024

    work page 2024

  70. [70]

    Explaining the black-box smoothly—a counterfactual approach,

    S. Singla, M. Eslami, B. Pollack, S. Wallace, and K. Batmanghe- lich, “Explaining the black-box smoothly—a counterfactual approach,” Medical Image Analysis, vol. 84, p. 102721, 2023

    work page 2023

  71. [71]

    Constrained decoding for secure code generation,

    Y. Fu, E. Baker, Y. Ding, and Y. Chen, “Constrained decoding for secure code generation,” arXiv preprint arXiv:2405.00218, 2024

    work page arXiv 2024

  72. [72]

    A framework to assess clinical safety and hallucination rates of llms for medical text summari- sation,

    E. Asgari, N. Montaña-Brown, M. Dubois, S. Khalil, J. Balloch, J. A. Yeung, and D. Pimenta, “A framework to assess clinical safety and hallucination rates of llms for medical text summari- sation,” npj Digital Medicine, vol. 8, no. 1, p. 274, 2025

    work page 2025

  73. [73]

    A framework for human evaluation of large language models in healthcare derived from literature review,

    T. Y. C. Tam, S. Sivarajkumar, S. Kapoor, A. V. Stol- yar, K. Polanska, K. R. McCarthy, H. Osterhoudt, X. Wu, S. Visweswaran, S. Fu et al., “A framework for human evaluation of large language models in healthcare derived from literature review,” NPJ digital medicine, vol. 7, no. 1, p. 258, 2024

    work page 2024

  74. [74]

    SAM 2: Segment Anything in Images and Videos

    N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson et al., “Sam 2: Segment anything in images and videos,” arXiv preprint arXiv:2408.00714, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  75. [75]

    Memory-SAM: Human-Prompt-Free Tongue Segmentation via Retrieval-to-Prompt

    J. Chae, L. Luo, X. Yuan, D. Yu, Z. Chen, L. Zhang, and P. Qin, “Memory-sam: Human-prompt-free tongue segmentation via retrieval-to-prompt,” arXiv preprint arXiv:2510.15849, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  76. [76]

    DINOv3

    O. Siméoni, H. V. Vo, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V. Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa et al., “Dinov3,” arXiv preprint arXiv:2508.10104, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  77. [77]

    The faiss library,

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.- E. Mazaré, M. Lomeli, L. Hosseini, and H. Jégou, “The faiss library,” IEEE Transactions on Big Data, 2025

    work page 2025

  78. [78]

    Qwen3 Technical Report

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv et al., “Qwen3 technical report,” arXiv preprint arXiv:2505.09388, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  79. [79]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen et al., “Lora: Low-rank adaptation of large language models. ” ICLR, vol. 1, no. 2, p. 3, 2022

    work page 2022

  80. [80]

    Efficient and robust ap- proximate nearest neighbor search using hierarchical navigable small world graphs,

    Y. A. Malkov and D. A. Yashunin, “Efficient and robust ap- proximate nearest neighbor search using hierarchical navigable small world graphs,” IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 4, pp. 824–836, 2018

    work page 2018

Showing first 80 references.