跳到正文
原文
Nature 健康护理主题·· 2026-10-16AI 评分38

基于知识蒸馏的轻量化生成式大语言模型用于中文放射学报告标准化质量控制

Knowledge distillation-based lightweight generative large language models for standardized quality control of Chinese radiology reports

AI 导读

研究人员开发了一种4B参数本地化部署语言模型,用于中文放射学报告的质量控制。教师模型Doubao-seed1.6-non-thinking达到最高宏精度0.94,最佳学生模型Qwen3-4B-Instruct-2507实现宏精度0.90,在真实临床场景中警报接受率为0.74。

正文

Scientific Reports (2026) Cite this article

We’re sharing this article early to provide faster access to peer-reviewed, accepted research. It is citable and carries a permanent DOI. This version is subject to further edits and will be replaced automatically by the final Version of Record. All legal disclaimers apply.

Abstract

Radiology reports, which document radiologists’ imaging observations and diagnostic conclusions from CT, X-ray and MRI scans, deliver core imaging evidence for clinical diagnosis and therapeutic planning. Inaccurate or disordered reports may lead to misdiagnosis and delayed intervention, posing risks to patient safety. Radiology reports are essential for clinical decision-making, yet Chinese radiology reports frequently exhibit logical inconsistencies, incomplete content, non-standardized descriptions, and formatting errors, while manual quality control remains time-consuming and inconsistent. Large teacher models can provide task-specific supervision but are difficult to deploy locally, whereas smaller models require adaptation to this clinical text task. We developed a teacher-supervised, locally deployable 4B language-model system for rule-based textual quality control of Chinese radiology reports. Among the three evaluated teacher APIs, Doubao-seed1.6-non-thinking achieved the highest macro-precision of 0.94, under the study’s precision-prioritized selection. Qwen3-4B-Instruct-2507 emerged as the optimal student model, achieving a macro-averaged precision of 0.90, recall of 0.49, and F1-score of 0.62. Under the deployment conditions of this study, it yielded an observed inference speed improvement of over 40-fold. Real-world clinical validation demonstrated that the lightweight model attains an alert acceptance rate of 0.74 in authentic clinical scenarios, suggesting it can serve as an adjunct for human radiologist review.

Abbreviations

PI-RADS:

Prostate imaging reporting and data system

BI-RADS:

Breast imaging reporting and data system

CT:

Computed tomography

MRI:

Magnetic resonance imaging

CTA:

Computed tomography angiography

MRA:

Magnetic resonance angiography

LLM:

Large language model

LoRA:

Low-rank adaptation

API:

Application programming interface

GPU:

Graphics processing unit

TP:

True positive

FP:

False positive

TN:

True negative

FN:

False negative

P:

Precision

R:

Recall

QC:

Quality control

Acknowledgements

This study was supported by The University Synergy Innovation Program of Anhui Province (Grant No. GXXT-2022-031) and the Key Program of the National Natural Science Foundation of China (Grant No. 62133004).

Author information

Authors and Affiliations

  1. Anhui University of Chinese Medicine, Hefei, China

    Duoning Jiang

  2. Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei, China

    Chuanfu Li

  3. Anhui Xunfei Imlincs Co., Ltd., Hefei, China

    Xiao Han & Ke Xu

  4. Hefei University of Technology, Hefei, China

    Xiao Han

  5. The First Affiliated Hospital of Anhui University of Chinese Medicine, Hefei, China

    Chuanfu Li

Authors

  1. Duoning Jiang
  2. Xiao Han
  3. Ke Xu
  4. Chuanfu Li

Corresponding author

Correspondence to Chuanfu Li.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Jiang, D., Han, X., Xu, K. et al. Knowledge distillation-based lightweight generative large language models for standardized quality control of Chinese radiology reports. Sci Rep (2026). https://doi.org/10.1038/s41598-026-76600-y

Download citation

  • Received: 28 August 2026

  • Accepted: 14 October 2026

  • Published: 16 October 2026

  • DOI: https://doi.org/10.1038/s41598-026-76600-y

Keywords

来源:Nature 健康护理主题 · nature.com