文献类型：期刊文献

英文题名：TriMRG: Toward Hallucination-Resistant Medical Image Report Generation via Triple Graph Semantic Reasoning and Rethinking

作者：Feng, Zehui[1]; Zhu, Dian[2]; Zhao, Jianan[3]; Zhao, Peiquan[4,5,6]; Ren, Jianing[7]; Han, Ting[8]

机构：[1] Shanghai Jiao Tong University, School of Design, Shanghai, 200240, China; [2] East China University of Science and Technology, School of Art Design and Media, Shanghai, 200240, China; [3] Donghua University, School of Design, Shanghai, 200240, China; [4] Shanghai Jiao Tong University, Xin Hua Hospital, School of Medicine, Shanghai, 200092, China; [5] International Ophthalmic Union, Task Force on the Prevention and treatment of diabetic eye diseases, China; [6] International Committee for the Classification of Retinopathy of Prematurity, China; [7] Shanghai Jiao Tong University, Xin Hua Hospital, School of Medicine, Department of Ophthalmology, Shanghai, 200092, China; [8] Shanghai Jiao Tong University, Institute of Medical Robotics, Shanghai, 201306, China

年份：2026

外文期刊名：IEEE Transactions on Circuits and Systems for Video Technology

收录：EI(收录号：20262921141767);Scopus(收录号：2-s2.0-105044817866)

语种：英文

外文关键词：Diseases - Iterative methods - Knowledge graph - Medical image processing

摘要：Medical image report generation (MRG) aims to translate complex visual diagnostic information into structured and coherent narratives, serving as a pivotal step in automated medical decision support. Existing methods fail to emulate the nuanced clinical reasoning and thinking processes of radiologists, modeling the multi-granularity semantic alignment across visual regions, disease concepts, and historical diagnostic knowledge, limiting their ability to generate context-aware and diagnostically faithful reports. In this work, we propose TriMRG, a novel LLM-based Triple Semantic Reasoning Generation framework that mimics the diagnostic workflow of human radiologists by jointly modeling region-level, disease-level, and memory-level semantics. Triple Semantic Graph Reasoning strategy constructs hierarchical semantic graphs with frequency-aware edge reweighting to avoid hallucination among memory, regions, and diseases; Generation Report Rethinking Match mechanism enhances global-local consistency between generated reports and retrieved exemplars, simulating clinicians’ iterative verification behavior. Extensive experiments on public MRG benchmarks and a real-world dataset demonstrate that TriMRG consistently outperforms existing state-of-the-art baselines in terms of clinical precision, semantic fidelity, and report coherence. In particular, on the MIMIC-CXR dataset, TriMRG achieves relative improvements of 1.4% in BLEU-4, 5.3% in METEOR, and 1.8% in RadGraph over the strongest competing methods, highlighting its effectiveness for reliable medical report generation. ? 1991-2012 IEEE.