人工智能赋能多模态影像学在炎症性肠病诊疗中的应用进展

The Application Progress of Artificial Intelligence Empowering Multimodal Imaging in the Diagnosis and Treatment of Inflammatory Bowel Disease

摘要: 炎症性肠病(IBD)的精准医疗高度依赖于影像学与病理学评估，然而，传统评估手段存在主观性强、重复性差以及信息割裂等局限。近些年来，人工智能(AI)，尤其是深度学习技术，为整合内镜、CT/MRI、肠道超声及组织病理学等多模态影像数据提供了全新范式。本文系统梳理近五年AI赋能多模态影像学在IBD中的应用进展。众多研究成果均表明，单模态AI模型在内镜Mayo评分、CTE图像识别及组织学炎症分级中已展现出接近专家的诊断性能(准确率80%~97%，AUC 0.74~0.99)；而多模态融合策略通过整合异构数据，在治疗反应预测(AUC达0.858)与组织学活动度评估(AUC达0.874)方面的准确性，显著优于单一模态方法。虽然多模态融合前景广阔，但实际中仍面临数据异质性、标注标准不一、模型泛化性不足及缺乏前瞻性验证等多种问题。未来我们应着力于构建标准化多模态数据库、发展可解释性大模型，并通过多学科协作来推动临床转化，以实现IBD从“经验决策”向“数据驱动的精准管理”的进阶。

Abstract: The precision medicine of inflammatory bowel disease (IBD) is highly dependent on imaging and pathological assessment. However, traditional assessment methods have limitations such as strong subjectivity, poor repeatability, and information fragmentation. In recent years, artificial intelligence (AI), especially deep learning technology, has provided a brand-new paradigm for integrating multimodal imaging data such as endoscopy, CT/MRI, intestinal ultrasound and histopathology. This article systematically reviews the application progress of AI-enabled multimodal imaging in IBD over the past five years. Numerous research results have shown that single-modal AI models have demonstrated diagnostic performance close to that of experts in endoscopic Mayo scoring, CTE image recognition, and histological inflammation grading (accuracy rate 80% to 97%, AUC 0.74 to 0.99). The multimodal fusion strategy, by integrating heterogeneous data, is significantly more accurate than the single-modal method in terms of treatment response prediction (AUC up to 0.858) and histological activity assessment (AUC up to 0.874). Although multimodal fusion has broad prospects, in practice, it still faces various problems such as data heterogeneity, inconsistent annotation standards, insufficient model generalization, and lack of prospective verification. In the future, we should focus on building standardized multimodal databases, developing large interpretable models, and promoting clinical transformation through multidisciplinary collaboration, so as to achieve the advancement of IBD from “experience-based decision-making” to “data-driven precise management”.