基于卷积神经网络预测早期胃癌浸润深度的 多模态AI系统开发与验证

Multimodal AI System for Predicting Early Gastric Cancer Invasion Depth Based on Convolutional Neural Networks: Development and Validation Study

摘要: 目的：开发并验证基于卷积神经网络的多模态AI系统，用于预测早期胃癌(EGC)浸润深度，辅助临床制定内镜下黏膜下剥离术(ESD)治疗决策。方法：回顾性收集2018年10月至2025年4月青岛大学附属医院经病理确诊的902例EGC患者的11,662张内镜图像，包括白光成像(WLI)、窄带成像(NBI)和靛胭脂染色图像。基于EfficientNet-B0架构构建CNN模型，通过内部验证集评估不同成像模式的诊断性能，并设计人机对比实验(图像层面和病例层面)比较AI系统与不同年资内镜医师的诊断效能。结果：内部验证集中，WLI模型表现最优，准确度87.6%、灵敏度89.1%、特异度85.2%、AUC 0.8836；WLI与NBI综合模型准确度85.3%，AUC 0.8634。人机对比实验中，AI系统在图像层面准确度86.65%、特异度87.47%，全面优于内镜医师；病例层面采用多数投票法整合信息后，AI准确度提升至92%，灵敏度93.75%，优势进一步扩大。结论：本研究成功开发的多模态AI系统能精准区分EGC黏膜内与黏膜下浸润，诊断效能显著优于不同年资内镜医师，病例层面整合多图信息后表现更优，具备辅助临床决策、促进诊断同质化的潜力。

Abstract: Background: Develop and test a multimodal AI system based on convolutional neural networks to predict the invasion depth of early gastric cancer (EGC) and assist clinical formulation of endoscopic submucosal stripping (ESD) treatment decisions. Methods: This study retrospectively collected a total of 11,662 endoscopic images of 902 patients diagnosed with early gastric cancer and adenocarcinoma by pathological diagnosis from October 2018 to April 2025 in the Affiliated Hospital of Qingdao University, including white light imaging (WLI), narrowband imaging (NBI) and indigo-carmine dye contrast images. The convolutional neural network (CNN) based on the EfficientNet-B0 architecture is used for model training and testing, and the model performance is evaluated through internal image verification and external human-computer comparison experiments (including image-based and case-based evaluation). Results: In the internal test set, the WLI model has the best performance, with accuracy of 87.6%, sensitivity of 89.1%, specificity of 85.2%, and AUC 0.8836; the accuracy of WLI + NBI models is 85.3%, and the AUC is 0.8634. In the human-computer comparison experiment, the AI system has an accuracy of 86.65% and a specificity of 87.47% at the image level, which is better than that of endoscopy doctors in terms of all aspects. After integrating information by majority voting at the case level, the accuracy of AI has increased to 92% and the sensitivity is 93.75%, and the advantage has been further expanded. Conclusions: Our study developed an AI system capable of predicting the invasion depth of EGC from multimodal endoscopic images, and the diagnostic performance superior to that of endoscopists. The model shows strong standardization and generalization capabilities, which has the potential to assist clinical decision-making and promote diagnosis homogenization.