LungCURE: Benchmarking Multimodal Real-World Clinical Reasoning for Precision Lung Cancer Diagnosis and Treatment

🏠 Homepage | 💻 GitHub | 📄 Paper (arXiv)

Overview

LungCURE is the first standardized multimodal benchmark designed for evaluating multimodal large language models (MLLMs) on precision lung cancer clinical decision support. It is constructed from 1,000 real-world, doctor-labeled clinical cases collected across more than ten hospitals in China, covering three clinically grounded evaluation tasks:

• Task 1 — TNM Staging: Predict the TNM stage from multimodal clinical materials.
• Task 2 — Treatment Recommendation: Generate guideline-compliant treatment plans given the ground-truth TNM stage.
• Task 3 — End-to-End Clinical Decision Support: Produce a complete diagnostic and treatment recommendation directly from raw clinical inputs, without any staging intervention.

All cases include multimodal clinical data such as imaging reports, pathology reports, clinical records, and genomic testing results, available in both Chinese (ZH) and English (EN).

---

Dataset Structure

LungCURE/
├── data/
│   ├── LC_patient_image/        # 1,000 Chinese/English multimodal clinical case PDFs (image format)
│   ├── LC_patient_text/        # 1,000 Chinese/English multimodal clinical case documents (text/OCR format)
└── ground_truth/
    ├── tnm-gt-ZH/       # Expert-annotated TNM staging ground truth (Chinese)
    ├── tnm-gt-EN/       # Expert-annotated TNM staging ground truth (English)
    ├── cds-gt-ZH/       # Clinical decision support ground truth (Chinese)
    └── cds-gt-EN/       # Clinical decision support ground truth (English)

Data Fields

data/ — Clinical Case Input (PDF)

Each case is stored as a single PDF file (e.g., LC_patient_0001.pdf) consolidating all available multimodal clinical materials for that patient, which may include:

Content	Description
Imaging report	CT / PET-CT radiology report with findings and diagnosis
Pathology report	Pathological examination and biopsy results
Clinical records	Patient history, ECOG status, lab results
Gene testing	Driver mutation / genomic profiling results (where available)

---

ground_truth/tnm-gt-*/ — TNM Staging Ground Truth (JSON)

Each file (e.g., LC_patient_0001.json) contains a structured JSON with detailed staging annotations:

Field	Description
q_tx	Whether the primary tumor is unassessable (Tx)
q_nx	Whether regional lymph nodes are unassessable (Nx)
q_t_size	T sub-category based on tumor size (e.g., T2a)
q_t2a / q_t2a_centr	Invasion characteristics supporting the T stage
q_n_stage	N stage(s) involved (e.g., N1, N2, N3)
q_n1_detail / q_n2_detail / q_n3_detail	Specific involved lymph node stations
q_m0	Whether the case is M0
q_m_detail	M sub-category (e.g., M1c)
q_m1c	Specific metastatic sites
q_m_uncertain	Uncertain/suspicious metastatic sites noted by annotators
q_difficulty	Annotator-assigned case difficulty (Easy / Medium / Difficult)

Example:

{
  "answers": {
    "q_nx": "No",
    "q_tx": "No",
    "q_t_size": "T2a (3 < T2a ≤ 4)",
    "q_t2a": ["T2a Centr"],
    "q_t2a_centr": ["Obstructive pneumonitis or atelectasis"],
    "q_n_stage": ["N1", "N2", "N3"],
    "q_n1_detail": ["Ipsilateral hilar"],
    "q_n2_detail": ["Ipsilateral mediastinal (stations 2, 3, 4, 5, and 6)"],
    "q_n3_detail": ["Supraclavicular lymph nodes (ipsilateral or contralateral)"],
    "q_m0": "No",
    "q_m_detail": ["M1c"],
    "q_m1c": ["Extrathoracic non-regional lymph nodes"],
    "q_m_uncertain": "Right cervical region",
    "q_difficulty": "Difficult"
  }
}

---

ground_truth/cds-gt-*/ — Clinical Decision Support Ground Truth (Markdown)

Each file (e.g., LC_patient_0001.md) contains a treatment plan provided by senior oncologists based on the patient's specific clinical situation. The content varies by case and may include one or more of the following:

• Diagnosis: Final staging conclusion, histological subtype, driver gene status, etc.
• Treatment strategy: Surgery, radiotherapy, chemotherapy, immunotherapy, targeted therapy, etc.
• Specific regimens: Preferred and alternative drug regimens following NCCN/CSCO guidelines
• Prognosis: Clinical assessment of the patient's prognosis
• Follow-up management: Post-treatment monitoring and maintenance therapy recommendations

Note: Not all fields are present in every case. The actual content depends on the specific clinical situation of each patient.

---

Dataset Statistics

Split	Language	Input Modality	# Cases
Test	Chinese (ZH)	Image (PDF)	1,000
Test	Chinese (ZH)	Text (OCR)	1,000
Test	English (EN)	Image (PDF)	1,000
Test	English (EN)	Text (OCR)	1,000

All data are used for evaluation only. LungCURE is a test-only benchmark.

---

Usage

Download the dataset

from huggingface_hub import hf_hub_download
import zipfile
 
# Download the zip file
zip_path = hf_hub_download(
    repo_id="Fine2378/LungCURE",
    filename="LungCURE.zip",
    repo_type="dataset",
    local_dir="./LungCURE"
)
 
# Extract the zip file
with zipfile.ZipFile(zip_path, "r") as zip_ref:
    zip_ref.extractall("./LungCURE")

---

Evaluation Tasks

Task 1: TNM Staging

Given multimodal clinical inputs, predict the TNM stage. A prediction is considered correct only if all three components (T, N, M) are simultaneously correct.

Task 2: Treatment Recommendation

Given multimodal clinical inputs and the ground-truth TNM stage, generate a guideline-compliant treatment recommendation. Evaluated on Precision and BERT-F1.

Task 3: End-to-End Clinical Decision Support

Given multimodal clinical inputs only (no staging provided), generate a complete clinical decision. This task reflects real-world deployment scenarios.

---

Annotation Protocol

All cases were annotated by senior oncologists following a two-stage protocol:

1. TNM Staging Annotation: Board-certified thoracic oncologists annotated each T, N, and M component with explicit clinical reasoning evidence, strictly following the AJCC 9th edition staging system.
2. Treatment Plan Generation: Reference treatment plans were generated strictly following NCCN and CSCO clinical guidelines, based on key clinical variables including histological subtype, driver gene status, PD-L1 expression, and performance status.

All data were fully de-identified and the study was approved by the Ethics Committee of Peking Union Medical College Hospital.

---

Citation

If you use LungCURE in your research, please cite:

@misc{hao2026lungcurebenchmarkingmultimodalrealworld,
      title={LungCURE: Benchmarking Multimodal Real-World Clinical Reasoning for Precision Lung Cancer Diagnosis and Treatment}, 
      author={Fangyu Hao and Jiayu Yang and Yifan Zhu and Zijun Yu and Qicen Wu and Wang Yunlong and Jiawei Li and Yulin Liu and Xu Zeng and Guanting Chen and Shihao Li and Zhonghong Ou and Meina Song and Mengyang Sun and Haoran Luo and Yu Shi and Yingyi Wang},
      year={2026},
      eprint={2604.06925},
      archivePrefix={arXiv},
      primaryClass={cs.MM},
      url={https://arxiv.org/abs/2604.06925}, 
}

---

License

This dataset is released under the MIT License.

---

Contact

For questions or issues, please open a GitHub issue at https://github.com/Joker-hfy/LungCURE or contact the corresponding authors.