基于多病种联合学习的脑肺疾病辅诊模型研究  

Research on the Auxiliary Diagnosis Model of Brain and Lung Diseases Based on Multi Disease Joint Learning

作　　者：叶荔姗 杨懿 YE Lishan;YANG Yi(Department of Automation,Tsinghua University,Beijing 100084,China)

机构地区：[1]清华大学自动化系,北京市100084 [2]厦门市卫生健康委规划发展与信息处,福建省厦门市361004 [3]北航杭州创新研究院,浙江省杭州市310020

出　　处：《中国卫生信息管理杂志》2025年第2期282-290,共9页Chinese Journal of Health Informatics and Management

基　　金：国家重大科技专项“慢性气道疾病肺功能智能诊断与云平台管理创新模式研究”(2023ZD0506300)。

摘　　要：目的针对现有医学影像辅助诊断系统局限于单一病种分析的问题,提出一种基于多病种联合学习的辅诊模型,通过跨病种特征融合与对比学习量化关联,提升诊断的全面性。方法整合脑部与肺部CT影像数据,构建多模态联合学习框架。设计跨病种特征融合机制,通过基于多模态数据的深度学习算法动态关联脑肺特征;引入对比学习驱动的关联分析,采用余弦相似度量化报告相关性;提出动态加权训练策略优化数据不平衡问题。基于21所医院6733例脑部与5657例患者胸肺部影像数据,对比单病种训练与联合训练的性能差异。结果联合训练模型在脑部测试集BLEU-1达0.632(单病种0.601),胸肺部测试集CIDEr-D提升至0.753(单病种0.704);跨病种报告匹配成功率显著高于随机水平(脑→肺55.6%,肺→脑62.2%)。结论基于多病种联合学习的脑肺疾病辅诊模型有效捕捉疾病间潜在关联,为临床提供了更全面的辅助诊断支持。Objective To address the limitation of existing medical image auxiliary diagnosis systems confined to single-disease analysis,a multi-disease joint learning-based auxiliary diagnostic model is proposed to enhance diagnostic comprehensiveness through cross-disease feature fusion and contrastive learning.Methods Brain and pulmonary CT imaging data were integrated to construct a multi-modal joint learning framework.A cross-disease feature fusion mechanism was designed to dynamically associate brain and pulmonary features via deep learning algorithm for multi-modal data.Correlation analysis driven by contrastive learning was introduced to quantify report correlations using cosine similarity,and a dynamic weighted training strategy was proposed to optimize data imbalance.Performance differences between single-disease training and joint training were compared using 6,733 brain and 5,657 pulmonary imaging cases from 21 Hospital.Results The joint training model achieved BLEU-1 of 0.632 on the brain test set(vs.0.601 for single-disease training)and CIDEr-D of 0.753 on the pulmonary test set(vs.0.704 for single-disease training).Cross-disease report matching success rates significantly exceeded random levels(brain→chest:55.6%,pulmonary→brain:62.2%).Conclusion Multi-disease joint learning effectively captures latent cross-disease correlations,providing comprehensive clinical diagnostic support.

关 键 词：多病种联合学习 医学影像报告生成 跨病种关联 智能辅助诊断 多模态数据 脑肺疾病 

分 类 号：R197.323[医药卫生—卫生事业管理] R319[医药卫生—公共卫生与预防医学]