摘要
目的通过开发一种基于卷积神经网络(convolutional neural network,CNN)的人工智能诊断模型,用于自动识别腺瘤和增生性息肉,以增加结直肠镜检查的统一性和客观性。方法回顾性收集南京医科大学第一附属医院内镜中心的白光内镜图像1796张,其中正常结直肠图像692张,增生性息肉608张,腺瘤496张。在我们的分类系统上提出双流网络(dual-stream network),包括原始流和检测流,原始流的输入是完整的预处理肠镜图像,以用来学习全局特征,检测流重点关注病灶的局部特征。我们将数据集分为训练集、测试集和验证集进行10次交叉验证,以评估该方法的有效性。结果CNN方法的正确率(accuracy)、准确率(precision)、召回率(recall)分别为96.9%、96.6%、96.7%,在正常肠镜图片、增生性息肉、腺瘤三类图像中的正确识别率分别为100%、95.1%和95.0%。结论CNN系统对结直肠息肉的识别具有较高的特异性和灵敏性,可以帮助临床内镜医师快速诊断和识别结直肠息肉的类型。这表明我们的方法在临床上能够对肠息肉病变进行有效、准确、可靠的诊断。
Objective To develope an artificial intelligence diagnosis model based on convolutional neural network(CNN)for automatic identification of adenomas and hyperplastic polyps.Methods 1796 white light endoscopy images were collected retrospectively,including 692 normal colorectal images,608 hyperplastic polyps and 496 adenomas.A dual-stream network is proposed in our classification system.The input of original stream was complete pre-processed enteroscopy image to focus on global features.Detected stream emphasizes on the local features of lesions.We divided the datasets into training set,test set,and verification set on 10-fold cross-validation to evaluate the proposed method.Results The accuracy of the proposed CNN method was 96.9%and the method’s precision,recall were 96.6%,96.7%.The classification accuracy of normal colorectal,hyperplastic polyps and adenomas were 100%,95.1%and 95.0%.Conclusion The CNN system has a high specificity and sensitivity for identification of colorectal polyps and can assist clinical endoscopists to quickly diagnose and identify the type of colorectal polyps.All of these indicate that our method enables an efficient,accurate and reliable esophageal lesion diagnosis in clinical.
作者
黄佩云
王凯妮
何晓璞
孙为豪
HUANG Peiyun;WANG Kaini;HE Xiaopu;SUN Weihao(Department of Geriatric Gastroenterology,the First Affiliated Hospital of Nanjing Medical University,Nanjing 210000;Biomedical Engineering of Southeast University,China)
出处
《胃肠病学和肝病学杂志》
CAS
2021年第3期336-340,共5页
Chinese Journal of Gastroenterology and Hepatology
基金
东南大学-南京医科大学合作研究项目(2242019K3DN03)。
关键词
结直肠癌
增生性息肉
腺瘤
卷积神经网络
人工智能
深度学习
Colorectal cancer
Hyperplastic polyps
Adenomas
Convolutional neural network
Artificial intelligence
Deep learning
作者简介
第一作者:黄佩云,硕士,研究方向:消化道肿瘤,人工智能。E-mail:peiyunhuang@njmu.edu.cn;通讯作者:孙为豪,博士,主任医师,教授,研究方向:消化道肿瘤,幽门螺旋杆菌感染。E-mail:sunweihao202010@163.com。
