胃肠道肿瘤人工智能诊疗系统：公共数据库建模与真实世界临床验证的整合

AI-Based Diagnostic and Therapeutic Systems for Gastrointestinal Tumors: Integrating Public Database Modeling with Real-World Clinical Validation

摘要: 胃肠道肿瘤具有发病隐匿、病理异质性强、治疗路径长等特点。临床实践中，单一内镜、影像、病理或组学信息往往难以覆盖筛查、诊断、分期、疗效预测及随访管理的连续需求。公共数据库为模型初步构建提供了规模化材料，真实世界数据则决定算法能否承受设备差异、病种谱变化与流程约束。本文围绕胃肠道肿瘤智能预测诊断系统的真实世界应用，综述TCGA、GEO、cBioPortal、TCIA等公共数据库的价值与边界，梳理内镜、放射影像、数字病理和多组学融合模型在胃癌、结直肠癌、食管癌等场景中的研究进展，并讨论从公共数据库建模到单中心回顾性验证、多中心外部验证、前瞻性静默试验和临床影响评价的转化路径。现有研究提示，多模态融合可在一定程度上提高对淋巴结转移、腹膜转移、微卫星不稳定性、病理完全缓解及预后的判别能力，但模型性能高度依赖数据质量、标注一致性、目标定义、分布偏移控制和可解释输出。未来研究应从“单点算法”转向“可审计的临床系统”，在数据治理、模型更新、伦理监管和医生–算法协同机制下形成可复现、可验证、可追踪的应用闭环。

Abstract: Gastrointestinal tumours are characterised by occult onset, marked biological heterogeneity and a long continuum of clinical management. Single-source information from endoscopy, radiology, pathology or omics is usually insufficient for screening, diagnosis, staging, treatment response prediction and follow-up. Public databases provide scalable materials for model exploration, whereas real-world data determine whether algorithms can tolerate variations in devices, disease spectrum and clinical workflow. This review summarises the value and limitations of public resources such as TCGA, GEO, cBioPortal and TCIA, discusses intelligent models based on endoscopy, radiomics, digital pathology and multi-omics, and proposes a translational pathway from public data mining to retrospective modelling, multicentre external validation, prospective silent testing and impact evaluation. The available evidence suggests that multimodal fusion may improve the identification of lymph node metastasis, peritoneal metastasis, microsatellite instability, pathological complete response and prognosis, but its reliability depends on data governance, annotation consistency, outcome definition, domain shift control and interpretable outputs. Future studies should move from isolated algorithms to auditable clinical systems.