Page 115 - 南京医科大学学报自然科学版
P. 115

第44卷第2期                           南京医科大学学报(自然科学版)
                  2024年2月                   Journal of Nanjing Medical University(Natural Sciences)     ·253 ·


               ·技术方法·

                一种超声心动图关键帧智能检测方法



                杜   悦 ,史中青 ,戚占如 ,曾子炀 ,郭冠军 ,姚                 静 ,罗守华 ,顾       宁  1,4*
                                               3
                                                                       3
                                                       2
                      1
                                                               2
                                       2
                              2
                南京医科大学生物医学工程与信息学院,江苏                   南京    211166;南京大学医学院附属鼓楼医院超声医学科,江苏                   南京
                1                                                  2.
                210008;东南大学生物科学与医学工程学院,江苏               南京 210096;南京大学医学院,江苏           南京 210093
                                                                    4
                      3
               [摘   要] 目的:探讨基于深度学习(deep learning,DL)的ResNet+VST模型在超声心动图关键帧智能检测方面的可行性。方
                法:选取南京大学医学院附属鼓楼医院超声医学科采集的 663 个动态图像含心尖二腔(apical two chambers,A2C)、心尖三腔
               (apical three chambers,A3C)与心尖四腔(apical four chambers,A4C)3类临床检查常用切面以及EchoNet⁃Dynamic公开数据集中
                280个A4C切面动态图像,分别建立南京鼓楼医院数据集与EchoNet⁃Dynamic⁃Tiny 数据集,各类别图像按4∶1方式划分为训练
                集和测试集,进行 ResNet+VST 模型的训练以及与多种关键帧检测模型的性能对比,验证 ResNet+VST 模型的先进性。结果:
                ResNet+VST模型能够更准确地检测心脏舒张末期(end⁃diastole,ED)与收缩末期(end⁃systole,ES)图像帧。在南京鼓楼医院数
                据集上,模型对 A2C、A3C 和 A4C 切面数据的 ED 预测帧差分别为 1.52±1.09、1.62±1.43、1.27±1.17,ES 预测帧差分别为 1.56±
                1.16、1.62±1.43、1.45±1.38;在EchoNet⁃Dynamic⁃Tiny 数据集上,模型对A4C切面数据的ED预测帧差为1.62±1.26,ES预测帧差
                为1.71±1.18,优于现有相关研究。此外,ResNet+VST 模型有良好的实时性表现,在南京鼓楼医院数据集与EchoNet⁃Dynamic⁃
                Tiny数据集上,基于GTX 3090Ti GPU对16帧的超声序列片段推理的平均耗时分别为21 ms与10 ms,优于以长短期记忆单元
               (long short⁃term memory,LSTM)进行时序建模的相关研究,基本满足临床即时处理的需求。结论:本研究提出的ResNet+VST
                模型在超声心动图关键帧检测的准确性、实时性方面,相较于现有研究有更出色的表现,该模型原则上可推广到任何超声切
                面,有辅助超声医师提升诊断效率的潜力。
               [关键词] 超声心动图;关键帧;深度学习
               [中图分类号] R445.1                   [文献标志码] A                       [文章编号] 1007⁃4368(2024)02⁃253⁃10
                doi:10.7655/NYDXBNSN230743


                An intelligent detection method of key frame in echocardiography

                                                            3
                                     2
                      1
                                                                                   2
                                                                         2
                                                                                                3
                                               2
                DU Yue ,SHI Zhongqing ,QI Zhanru ,ZENG Ziyang ,GUO Guanjun ,YAO Jing ,LUO Shouhua ,GU Ning  1,4*
                                                                                                    2
                1 School of Biomedical Engineering and Informatics,Nanjing Medical University,Nanjing 211166;Department of
                                                                                                        3
                Ultrasound,Affiliated Drum Tower Hospital,Medical School,Nanjing University,Nanjing 210008;School of
                Biological Sciences and Medical Engineering,Southeast University,Nanjing 210096;Medical School of Nanjing
                                                                                         4
                University,Nanjing 210093,China
               [Abstract] Objective:To explore the feasibility of using ResNet+VST model based on deep learning(DL)for intelligent detection of
                key frames in echocardiography. Methods:A total of 663 dynamic images including apical two chambers(A2C),apical three chambers
               (A3C),and apical four chambers(A4C),which are commonly used clinical examination views,were collected from the Department of
                Ultrasound Medicine at Drum Tower Hospital,Nanjing University Medical School. Additionally,280 dynamic A4C images from the
                EchoNet ⁃ Dynamic public dataset were selected. Two datasets were established:the Nanjing Drum Tower Hospital dataset and the
                EchoNet⁃Dynamic⁃Tiny dataset. The images in each category were divided into training set and testing sets in a 4:1 ratio. The ResNet+
                VST model was trained and its performance was compared with other key frame detection models to verify the its superiority. Results:
                The ResNet+VST model can detect the end⁃diastolic(ED)and end⁃systolic(ES)image frames of the heart more accurately. On the
                Nanjing Drum Tower Hospital dataset,the model achieved ED frame prediction differences of 1.52±1.09,1.62±1.43,and 1.27±1.17 for
                A2C,A3C,and A4C views,respectively,and ES frame prediction differences of 1.56±1.16,1.62±1.43,and 1.45±1.38,respectively. On

               [基金项目] 江苏省重点研发计划(BE2022828);江苏省前沿引领技术基础研究专项(BK20222002)
                ∗
                通信作者(Corresponding author),E⁃mail: guning@nju.edu.cn
   110   111   112   113   114   115   116   117   118   119   120