关键词: 医学报告生成, 智能辅助决策, 多任务学习, 多尺度融合, 注意力机制

Abstract: With the rapid development of next-generation information technologies and their deep integration into the healthcare domain, cross-domain multimodal data analysis and intelligent decision support in the Internet of Healthcare Systems(IHS) environmentms have created new opportunities to improve the efficiency and quality of healthcare services. These approaches have been widely applied across multiple stages of disease prevention, diagnosis, treatment, and rehabilitation. In clinical practice, radiologists and imaging specialists are required to generate medical examination reports based on imaging findings to support clinical decision-making. However, medical imaging examinations often involve high visual similarity across different anatomical sites, and report writing requires the integration of clinical expertise to accomplish multiple tightly coupled tasks, including anatomical site classification, lesion region detection, and the use of standardized medical terminology. Under time constraints, less experienced physicians in primary healthcare settings often struggle to cope with the high-intensity and continuous workload of report writing. Automated medical report generation enables the direct transformation from medical images to examination reports, effectively reducing physicians’ workload while enhancing the diagnostic capacity and service quality of primary healthcare institutions. To address the challenges of insufficient discriminability of visual representations, multi-task procedural complexity, and the complex mapping between visual semantics and medical terminology, this study proposes an automated medical report generation method that integrates clinical diagnostic reasoning. By mining scale-level associations between multi-scale visual features and textual representations, the proposed method jointly performs anatomical site classification, lesion region detection, and report generation. Using endoscopic imaging examinations as a case study, extensive experiments were conducted on real-world clinical data from a tertiary hospital as well as large-scale public datasets to validate the effectiveness of the proposed approach. In clinical comparative experiments, professional clinicians evaluated 1,439 generated reports, and the results show that the qualified rate reached 78.70%, further highlighting the potential value of the proposed method for clinical decision support applications.

Key words: Medical report generation, Intelligent assistant decision-making, Multi-task learning, Multi-scale fusion, Attention mechanism