基于卷积神经网络的早期胃癌识别与检测研究

China is a high incidence area of gastric cancer, which accounts for 40% of the world for both new incidence and mortality of gastric cancer. Even though clinical studies have shown that most of the patients with early gastric cancer (EGC) can be cured under endoscopy, unfortunately, due to weak features of gastric images and varieties of EGC, the detection rate of EGC among all gastric cancer is only about 5% in china, which is far less than 50% in developed countries. The technology of EGC recognition and real-time detection based on deep learning can assist gastroscopists to screen out lesions effectively in gastroscopic operation, which can improve EGC detection rate in China and has absolutely prospective for scientific research and clinical applications. For this project, we will conduct innovative explorations from the following three aspects. (1) Set up a deep learning data platform for gastroscopic images. By fusing multi-modal data, including white light endoscopy(WLE), ME-NBI endoscopy and pathology data, we will establish a standard EGC dataset with manual annotation. (2) Research on the architecture of EGC recognition and detection convolutional neural networks (CNN). To improve the model’s performance, we will mainly focus on multi-modal data fusion, reinforce learning and optimization of model’s super parameters. (3) We will implement a real-time EGC detection system for gastroscopic video via reducing the number of CNN parameters, optimizing the architecture of networks, and selection of informative key frame, etc., which will provide gastroscopists with EGC decision support.

我国属于胃癌高发区域，每年胃癌的新增患者与死亡患者均占全球的40%。尽管大部分的早期胃癌（EGC）在内镜下即可获得根治性治疗，然而由于早期胃癌的特征不明显，种类繁多，我国常规胃镜检查检出的早期胃癌仅占全部胃癌的5%左右，远低于发达国家的50%。基于深度学习的EGC实时识别和检测技术能辅助医生高效筛查胃镜手术中的病灶，对提高EGC检出率具有重要的科研价值和应用前景。本项目将从以下三个方面进行创新性探索：（1）构建胃镜影像数据平台，融合白光胃镜、ME-NBI胃镜和病理等多模数据，建立具有人工标注的EGC标准样本数据集；（2）研究面向EGC识别和检测的卷积神经网络（CNN）模型，从多模态数据融合、强化学习和模型超参数优化等三个方面提升模型性能；（3）通过优化网络结构、减小CNN参数、视频图像关键帧选取等方法实现对胃镜视频实时检测，进而研制EGC实时辅助检测系统，辅助医生提高EGC的检出能力。

本项目在胃镜图像下构建早期胃癌（EGC）分类、识别和检测标准图像库的基础上，使用基于卷积神经网络的深度学习方法，研究和设计了相应的深度学习分类与检测模型，实现了面向胃癌早期诊断的多模态图像引导内镜系统。具体包括：.1）构建了基于Web的胃镜图像大数据平台。平台包含多中心消化内镜影像数据收集、数据标注、病灶和解剖标签管理以及病例管理的功能。内窥镜下白光和NBI等不同模态影像数量达41.7万余张，其中包含标注数据6万余张，且所有标注数据均经过了多专家并结合金标准病理结果的交叉验证。该数据平台的标注数据可按需导出用作模型构建和研究。.2）项目针对从正常胃粘膜逐渐演化的癌前疾病、癌前病变直到早期胃癌的各项病灶的变化过程，分别构建了标注数据集和智能识别模型，实现了完整覆盖胃部内窥镜下多种病灶的深度学习识别和检测模型的多中心训练和验证。.3）项目实现了功能性胃镜下图像智能感知模型的构建，包含：解剖位置识别模型、影像模态识别模型、视频有效帧识别模型等深度学习模型的训练和验证，为构建全自动智能辅助诊断系统的实现提供了算法基础。.4）综合上述建立的病灶模型和功能性模型，项目构建了基于深度学习的EGC辅助检测与诊断系统，能够在胃镜检查过程中实现检查质量控制与引导并辅助医生发现与诊断早期胃癌。目前我们已在浙江大学附属邵逸夫医院进行了300例以上的临床实验评估和验证，结果表明该系统达到甚至超越了临床医生的水平。该系统为解决内窥镜检查过程中面对的检查质量、胃癌相关的病灶漏诊和误诊等问题提供端到端的实时智能辅助诊断解决方案。