Scholarlink

Automated medical report generation is a critical task in intelligent healthcare, aiming to reduce physician workload while improving diagnostic efficiency and reporting consistency. In nasal endoscopy, variations in anatomical morphology, imaging angles, and detail representation often introduce noise into ground-truth reports, such as missed findings or incomplete clinical information. These issues primarily arise from anatomical regional differences and complex imaging conditions rather than human error. To address this challenge, we propose a Set-to-One Structured Captioning paradigm to generate unified and clinically consistent reports from diverse nasal endoscopy frames, effectively mitigating the noise caused by incompleteness and imprecision in human annotations. The proposed SetCap framework comprises two key modules: (1) Spatiotemporal Context-Aware Captioning (SCAC), which models anatomical transitions across frames via an anatomy-aware sliding window to capture local temporal context and ensure spatial consistency and temporal coherence; and (2) Medical Knowledge Retrieval and Refinement (MKRR), which leverages medical knowledge bases to retrieve and filter relevant entries, injecting domain knowledge to enhance semantic representation, standardize terminology, and reduce semantic ambiguity. Experiments conducted on three datasets: IRA-HUT-NA, IU-Xray, and MIMIC-CXR demonstrate that SetCap significantly outperforms state-of-the-art methods, achieving improvements of +12.3% in BLEU-4 and +15.2% in F1. Source code is available at: https://anonymous.4open.science/r/SetCap-AD82 .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Set-to-One Structured Captioning for Heterogeneous Nasal Image Collections via Spatio-Temporal Modeling and Clinical Knowledge Integration

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xinpan Yuan,
  • Jianuo Ju,
  • Liujie Hua,
  • Mingzhu Huang

摘要

Automated medical report generation is a critical task in intelligent healthcare, aiming to reduce physician workload while improving diagnostic efficiency and reporting consistency. In nasal endoscopy, variations in anatomical morphology, imaging angles, and detail representation often introduce noise into ground-truth reports, such as missed findings or incomplete clinical information. These issues primarily arise from anatomical regional differences and complex imaging conditions rather than human error. To address this challenge, we propose a Set-to-One Structured Captioning paradigm to generate unified and clinically consistent reports from diverse nasal endoscopy frames, effectively mitigating the noise caused by incompleteness and imprecision in human annotations. The proposed SetCap framework comprises two key modules: (1) Spatiotemporal Context-Aware Captioning (SCAC), which models anatomical transitions across frames via an anatomy-aware sliding window to capture local temporal context and ensure spatial consistency and temporal coherence; and (2) Medical Knowledge Retrieval and Refinement (MKRR), which leverages medical knowledge bases to retrieve and filter relevant entries, injecting domain knowledge to enhance semantic representation, standardize terminology, and reduce semantic ambiguity. Experiments conducted on three datasets: IRA-HUT-NA, IU-Xray, and MIMIC-CXR demonstrate that SetCap significantly outperforms state-of-the-art methods, achieving improvements of +12.3% in BLEU-4 and +15.2% in F1. Source code is available at: https://anonymous.4open.science/r/SetCap-AD82 .