Page 137 - 南京医科大学自然版
P. 137

第46卷第7期                           南京医科大学学报(自然科学版)
                  2026年7月                   Journal of Nanjing Medical University(Natural Sciences)     ·1083 ·


               ·综 述·

                从结构化分级到智能分级:人工智能与ACR⁃RADS融合的研究

                进展



                黄润楸 ,徐飞佳 ,刘育宏 ,冯钰淇 ,曾锦辉 ,梁定邦 ,周                      淳 ,汪    洋  1,2*
                                       3
                      1
                              2
                                               4
                                                                       4
                                                       3
                                                               3
                南方医科大学珠江医院 放射科,广东              广州   510280;同济大学附属第十人民医院 放射科,上海               200072;南方医科大学第
                1                                         2                                        3
                二临床医院,第一临床医学院,广东 广州                510515
                          4
               [摘   要] 影像报告与数据系统(reporting and data system,RADS)是美国放射学会(American College of Radiology,ACR)旨在
                降低影像判读主观性、提升一致性的结构化风险分层框架。虽已广泛应用,但在真实世界实践中仍面临评分一致性受限、特征
                量化不足及跨中心可重复性差等方法学挑战。近年来,人工智能(artificial intelligence,AI)尤其是大语言模型(large language
                model,LLM)的发展为解决上述局限提供了新路径。文章系统梳理了主要RADS体系的结构特征及实践局限,综述了AI在病
                灶识别、风险再分层及流程规范化等方面的实证进展。现有证据显示,AI更适宜作为RADS的“增强层”,以提升其客观性与重
                复性。文章重点探讨了LLM在语义理解、自动推理及质控中的潜力,并从监管视角展望了人机协作的演进方向。该趋势对推
                动我国影像诊断同质化、提升基层诊疗水平及构建智能化监管体系具有重要指导意义。
               [关键词] 影像报告与数据系统;人工智能;结构化影像报告;风险分层
               [中图分类号] TP18                    [文献标志码] A                       [文章编号] 1007⁃4368(2026)07⁃1083⁃09
                doi:10.7655/NYDXBNSN251492



                From structured reporting to intelligent grading:research advances in the integration of
                artificial intelligence and ACR⁃RADS
                              1         2           3          4            3               3           4
                HUANG Runqiu ,XU Feijia ,LIU Yuhong ,FENG Yuqi ,ZENG Jinhui ,LIANG Dingbang ,ZHOU Chun ,WANG
                Yang 1,2*
                                                                                                    2
                1 Department of Radiology,Zhujiang Hospital,Southern Medical University,Guangdong 510280;Department of
                                                                                 3
                Radiology,Tenth People’s Hospital of Tongji University,Shanghai 200072;Second School of Clinical Medicine,
                First School of Clinincal Medicine,Southern Medical University,Guangdong 510515,China
                4

               [Abstract] The reporting and data systems(RADS),established by the American College of Radiology(ACR),serve as structured
                risk⁃stratification frameworks designed to mitigate subjectivity and enhance consistency in radiological interpretation. Despite their
                widespread adoption,RADS still face methodological challenges in real⁃world practice,including limited inter⁃observer consistency,
                insufficient feature quantification,and suboptimal cross⁃center reproducibility. In recent years,the evolution of artificial intelligence
               (AI),particularly large language models(LLMs),has offered novel pathways to address these limitations. This paper systematically
                reviews the structural characteristics and practical constraints of major RADS frameworks and synthesizes empirical progress regarding
                AI in lesion identification,risk re⁃stratification,and workflow standardization. Current evidence suggests that AI is best positioned as
                an“augmentative layer”for RADS to bolster objectivity and reproducibility. Furthermore,this article explores the potential of LLMs in
                semantic understanding,automated reasoning,and quality control,while projecting the evolution of human⁃computer collaboration from
                a regulatory perspective. This trend holds significant implications for promoting the homogenization of diagnostic imaging,empowering
                primary healthcare services,and establishing intelligent regulatory systems in China.
               [Key words] reporting and data system;artificial intelligence;structured imaging reporting;risk stratification
                                                                            [J Nanjing Med Univ,2026,46(07):1083⁃1091]
               [基金项目] 广东省自然科学基金项目(2023A1515012499,2025A1515010318);广州市科技计划重点项目(2024B03J0781);
                广东省医学装备学会科研基金项目(YZXH2025KT10);广东省基础与应用基础研究基金(2024A1515220081)
                通信作者(Corresponding author),E⁃mail:wangyang98289@smu.edu.cn(ORCID:0000⁃0003⁃1588⁃3208)
                ∗
   132   133   134   135   136   137   138   139   140   141   142