基于不同卷积神经网络构建小肠多病变自动检测的人工智能辅助系统  

Construction of an artificial intelligence-assisted system for automatic detection of multiple lesions in the small intestine based on various convolutional neural networks

作　　者：陈健[1] 王甘红 张子豪 夏开建 戴建军[1] 徐晓丹 Chen Jian;Wang Ganhong;Zhang Zihao;Xia Kaijian;Dai Jianjun;Xu Xiaodan(Department of Gastroenterology,Changshu Hospital Affiliated to Soochow University,Suzhou 215500,Jiangsu,China;Department of Gastroenterology,Changshu Traditional Chinese Medicine Hospital,Suzhou 215500,Jiangsu,China;Shanghai Hao Brothers Educational Technology Co.,Ltd.,Shanghai 200434,China;Changshu Key Laboratory of Medical Artificial Intelligence and Big Data,Suzhou 215500,Jiangsu,China)

机构地区：[1]常熟市第一人民医院消化内科,江苏苏州215500 [2]常熟市中医院消化内科,江苏苏州215500 [3]上海豪兄教育科技有限公司,上海200434 [4]常熟市医学人工智能与大数据重点实验室,江苏苏州215500

出　　处：《兰州大学学报（医学版）》2024年第9期23-29,共7页Journal of Lanzhou University（Medical Sciences）

基　　金：苏州市第二十三批科技发展计划资助项目(SLT2023006);常熟市医学人工智能与大数据重点实验室能力提升资助项目(CYZ202301);常熟市医药卫生科技计划资助项目(CSWS202316)。

摘　　要：目的基于不同卷积神经网络开发人工智能辅助系统,用于自动检测小肠胶囊内镜(CE)拍摄的11类小肠病变图像,提高诊断的效率、准确性和客观性。方法收集来自3个不同医学中心,使用3种不同品牌设备拍摄的CE图像,构建图像数据集,用于不同卷积神经网络模型的训练和测试,共含13683张图像和15117个注释标签。模型性能评估指标包括平均精度、准确率、敏感性、特异性、假阳性率、检测速度。结果构建了2种YOLO模型和2种RTMDet模型,在包含2729张CE图像(4801注释标签)的测试集上,RTMDet_m模型取得了最佳的mAP50(82.58%),但也展现出最慢的延迟时间(47.28帧/s)。模型达到了82.76%的整体敏感性和95.91%整体准确率;在具体类别的推理中,敏感性最高的类别是“出血”,而最低的类别是“黏膜下肿瘤”。结论使用混合品牌CE图像开发的人工智能模型能够快速准确地检测与分类11种小肠病变,在帮助医师提升CE诊断效率和准确性方面展现出很好的临床应用潜力。Objective To develop an artificial intelligence-assisted system based on various convolutional neural networks for automatically detecting 11 types of small intestinal lesion images captured by capsule endoscopy(CE),aiming to enhance the efficiency,accuracy,and objectivity of diagnoses.Methods Images were collected from three medical centers across different countries using three different brands of CE equipment,building an image dataset for training and testing various convolutional neural network models,totaling 13683 images and 15117 annotated labels.The model performance metrics included mean precision,accuracy,sensitivity,specificity,false-positive rate,and detection speed.Results Two YOLO models and two RTMDet models were developed in this study.In a test set containing 2729 CE images(4801 annotated labels),the RTMDet_m model achieved the best mAP50(82.58%),but also showed the slowest latency at 47.28 frames per second.Additionally,this model achieved an overall sensitivity of 82.76%and an accuracy of 95.91%;in category-specific inference,the highest sensitivity was observed in the"bleeding"category,while the lowest was for"submucosal tumor".Conclusion The artificial intelligence models developed using mixed-brand CE images can rapidly and accurately detect and classify 11 types of small intestinal lesions,demonstrating significant clinical application potentials in enhancing the efficiency and accuracy of CE diagnoses.

关 键 词：小肠病变 卷积神经网络 人工智能 胶囊内镜 目标检测 

分 类 号：R574.5[医药卫生—消化系统] R318.14[医药卫生—内科学]