DAMEC: Disease-Aware Multi-Expert Consensus Framework for Study-Level Radiology Report Generation

CIKM 2026 Oral Presentation

This repository contains the official implementation of DAMEC: Disease-Aware Multi-Expert Consensus Framework for Study-Level Radiology Report Generation.

Junyeong Maeng, Eunsong Kang, and Heung-Il Suk, “DAMEC: Disease-Aware Multi-Expert Consensus Framework for Study-Level Radiology Report Generation,” Proceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM 2026), Rome, Italy, November 7–11, 2026. [Paper]

DAMEC formulates radiology report generation as a study-level, clinical-context generation task. A variable number of chest X-ray images within a study are analyzed by four heterogeneous experts—ConvNeXt, RAD-DINO, PriorRG, and MedGemma—and their disease-level predictions are integrated through a trainable disease-aware consensus module.

The resulting structured Clinical Findings (CF) descriptor serves as an explicit clinical anchor for clinical-context retrieval, report generation, and post-generation validation.

Overall Framework

Overview of the proposed DAMEC framework.

DAMEC consists of three main components:

• Study-level Multi-Expert Consensus: Multiple heterogeneous experts analyze the available chest X-ray views, and their disease-level predictions are aggregated into study-level clinical findings.
• Clinical Findings Representation: The consensus results are organized into a structured CF descriptor containing disease states and clinically relevant information.
• CF-Grounded Report Generation and Validation: The CF descriptor guides retrieval of clinically similar cases, report generation, and post-generation validation.

Installation

Create the environment and install the required packages:

conda create -n damec python=3.11 -y
conda activate damec
pip install -r requirements.txt

Please refer to INSTALL.md for model preparation, checkpoints, and environment configuration.

Usage

Configure the required dataset paths, model checkpoints, and LLM endpoints before running DAMEC.

# Precompute expert predictions
python scripts/precompute_experts.py --split test --config configs/local.yaml

# Train the consensus module
cd training
python train_consensus.py --config configs/consensus_default.yaml

# Run DAMEC
cd ..
python run.py --split test --config configs/local.yaml

Additional configuration and preprocessing instructions are provided in INSTALL.md.

Datasets

DAMEC is evaluated on four chest X-ray report generation benchmarks:

• MIMIC-CXR
• MIMIC-ABN
• Two-view CXR
• CheXpert Plus

Please follow the official data access and licensing policies of each dataset.

Citation

If you find this work useful, please cite:

@inproceedings{maeng2026damec,
  title     = {DAMEC: Disease-Aware Multi-Expert Consensus Framework for Study-Level Radiology Report Generation},
  author    = {Maeng, Junyeong and Kang, Eunsong and Suk, Heung-Il},
  booktitle = {Proceedings of the 35th ACM International Conference on Information and Knowledge Management},
  year      = {2026},
  doi       = {10.1145/3799682.3841082}
}

License

This project is released under the MIT License. See LICENSE for details.