[Construction of a Multimodal Forensic Pathology Database and Implementation of Model-Augmented Retrieval].

内容摘要

To construct a multimodal forensic pathology database based on artificial intelligence (AI) technology and explore methods for integrating pre-trained models into multimodal databases. A hybrid storage architecture consisting of MySQL, Redis and OSS was employed to manage desensitized multimodal data. Data entry was optimized using optical character recognition (OCR) and natural language processing (NLP) technologies. OCR performance was evaluated and optimized using cosine distance, character error rate (CER), and word error rate (WER) to meet practical operational requirements. An intelligent retrieval model was developed using the ChatGLM3-6B model and retrieval-augmented generation (RAG) technology. Model performance was evaluated using ranking metrics including Precision@K, Recall@K, discounted cumulative gain (DCG), and normalized discounted cumulative gain (NDCG). The database demonstrated satisfactory baseline performance. The response times were maintained within 150 ms for single-query condition and within 2 s for multiple-query conditions. The average disk read throughput reached 950 MB/s. In concurrent performance tests, the database achieved a maximum throughput of 1 200 queries per second (QPS), meeting multimodal data management demands. OCR evaluation showed high recognition accuracy; for high-quality documents, the cosine distance, CER, and WER achieved 0.02, 1.5%, and 3.2%, respectively. Intelligent retrieval results indicated that Precision@K remained consistently high (0.69-1.00), while NDCG values remained above 0.87 for all evaluations. When K=100, the NDCG surpassed 0.95 for all queries, meeting expected performance requirements. The multimodal forensic pathology database constructed in this study demonstrates good stability and operational efficiency and can meet the requirements of routine forensic practice for multimodal data storage, management, and analysis. The intelligent retrieval capabilities, based on pre-trained large language models (LLMs), can be applied to conversational information retrieval from forensic reports and related documents, providing a novel approach to the management and analysis of multimodal databases. 目的: 基于人工智能技术构建法医病理多模态数据库系统，并探索预训练模型在多模态数据库中集成的应用方法。方法: 系统采用MySQL、Redis和OSS的混合式存储架构管理经脱敏处理后的多模态数据。通过光学字符识别（optical character recognition，OCR）和自然语言处理（natural language processing，NLP）技术对数据录入过程进行优化，根据余弦距离、字符错误率和词错误率评估并优化OCR性能以满足实际业务需求。基于ChatGLM3-6B模型和检索增强生成（retrieval-augmented generation，RAG）技术构建智能检索模型，并采用精确率（Precision@K）、召回率（Recall@K）、折损累积增益（discounted cumulative gain，DCG）和归一化折损累积增益（normalized discounted cumulative gain，NDCG）等排名指标评估模型性能。结果: 数据库基准性能良好，单个查询条件的响应时间控制在150 ms内，多个查询条件在2 s内，平均磁盘读取吞吐量为950 MB/s，并发性能测试中数据库吞吐量最高可达1 200每秒查询率（queries per second，QPS），能够承担多模态数据管理的需求。OCR结果显示，系统识别精度较高;在文档质量较高时，余弦距离、字符错误率和词错误率分别达到0.02、1.5%、3.2%。智能检索结果显示，Precision@K维持在较高水平（0.69~1.00），NDCG值均保持在0.87以上;K=100时，所有查询的NDCG均超过0.95，达到预期要求。结论: 所构建的法医病理多模态数据库具有较好的稳定性和运行效率，能够满足法医日常工作中对多模态数据存储和管理分析的需求。基于预训练的大语言模型的智能检索功能可用于鉴定文书的对话式信息检索，为多模态数据库的管理分析提供新的途径。.