基于迁移学习和卷积神经网络多网络颈动脉图像分类检测研究
Research on Multi Network Carotid Artery Image Classification and Detection Based on Transfer Learning and CNN
隋小瑜 1韩晶 2曹艳坤 1米加 3宋延云 4王剑磊 5王春 5刘治1
作者信息
- 1. 山东大学信息科学与工程学院,山东 青岛 266237
- 2. 中国矿业大学信息与控制工程学院,江苏徐州 221116
- 3. 山东省立第三医院特检科,济南 250031
- 4. 山东省运动康复研究中心,济南 250102
- 5. 山东大学光学高等研究中心,山东青岛 266237
- 折叠
摘要
颈动脉超声检查是用于诊断斑块主要且方便的手段,从超声图像中准确获取关于斑块信息对于进一步临床诊断至关重要.由于超声仪器存在噪声的干扰以及人工技术操作差异,使得显示的切面图像不清晰、不标准,从而容易导致斑块的误检漏检.本研究提出一种基于迁移学习和CNN的深度学习算法多网络组合实现对颈动脉斑块更准确识别的研究.首先选取颈动脉清晰和模糊的共计2 591张纵切面超声图像通过ResNet网络进行管腔质量分类控制;进行管腔分类后,选取清晰的软斑和硬斑信息的纵切面图像共计1 114张,通过基于迁移学习的RetinaNet网络对其进行颈动脉管腔和斑块的分类检测并使用Faster R-CNN和SSD网络进行对比实验.对于管腔分类网络,测试集分类准确率达到93％;对于管腔和斑块分类检测网络,使用113张测试集图像得到管腔检测的平均精度在交并比(IOU)值为0.5时达到1,在IOU值为0.75时达到0.988,平均召回率达到0.869,均高于Faster R-CNN和SSD网络;得到硬斑和软斑检测的平均精度在IOU=0.5时分别达到0.899和0.855,平均召回率分别达到0.609和0.578.在进行颈动脉斑块分类识别前,进行颈动脉管腔图像的质量分类控制,能够有效避免图像不清晰带来的斑块误检,提高斑块检测的正确性,并对后续进行颈动脉三维重建具有重要意义.
Abstract
Carotid ultrasound is a main and convenient method for plaque diagnosis.Therefore,it is very important to obtain accurate information about plaque from ultrasound images for further clinical diagnosis.Due to the noise interference of ultrasonic machine and the difference of manual technical operation,the displayed section image is not clear and standard,which is easy to lead to false detection or missed detection of the plaque.In this work,a deep learning algorithm based on migration learning and CNN was proposed to realize the research of more accurate identification of carotid plaque.Firstly,2591 longitudinal ultrasound images with clear and fuzzy carotid artery were selected to classify and control the lumen quality through ResNet network;After the lumen classification,1114 longitudinal images with clear soft and hard plaque information were selected.The carotid lumen and plaque were classified and detected by RetinaNet network based on migration learning,and the comparative experiment was carried out by using Faster R-CNN and SSD network.For the lumen classification network,the classification accuracy of the test set was 93％.For the lumen and plaque classification detection network,113 test set images were used to obtain the average accuracy of lumen detection,which reached 1 when the intersection union ratio(IOU)value was 0.5,0.988 when the IOU value was 0.75,0.838 when the IOU value was 0.5:0.95,and the average recall rate reached 0.869,which were higher than those of Faster R-CNN and SSD networks;The average accuracy of hard plaque and soft plaque detection was 0.899 and 0.855 when IOU=0.5,and the average recall was 0.609 and 0.578 respectively.Before the classification and recognition of carotid plaque,the quality classification control of carotid lumen image can effectively avoid the false detection of plaque caused by unclear image,improve the correctness of plaque detection,and is of great significance for the follow-up three-dimensional reconstruction of carotid artery.
关键词
颈动脉超声图像/管腔质量分类控制/管腔检测/斑块分类识别Key words
carotid ultrasound images/lumen quality classification control/lumen detection/plaque classification and recognition引用本文复制引用
基金项目
山东省自然科学基金(ZR2019ZD05)
山东自然科学基金智能计算联合基金(ZR2020LZH013)
计算机体系结构国家重点实验室开放项目(CARCHA202002)
出版年
2024
NETL
NSTL
万方数据