基于深度学习的B超图像肝癌病灶自动定位  

Automatic localization of liver cancer in B-ultrasound images based on deep learning

作　　者：黄良汇 许祥丛 谭海曙[1,3] 韩定安 王雪花[1,3] HUANG Liang-hui;XU Xiang-cong;TAN Hai-shu;HAN Ding-an;WANG Xue-hua(School of Physics and Optoelectronic Engineering,Foshan University,Foshan 528225,China;School of Mechatronics Engineering and Automation,Foshan University,Foshan 528225,China;Guangdong-Hong Kong-Macao Joint Laboratory for Intelligent Micro-Nano Optoelectronic Technology,Foshan University,Foshan 528225,China)

机构地区：[1]佛山科学技术学院物理与光电工程学院,广东佛山528225 [2]佛山科学技术学院机电工程与自动化学院,广东佛山528225 [3]粤港澳智能微纳光电技术联合实验室,广东佛山528225

出　　处：《佛山科学技术学院学报（自然科学版）》2022年第6期21-27,共7页Journal of Foshan University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61805038,62075042);粤港澳智能微纳光电技术联合实验室研究基金(2020B1212030010)。

摘　　要：利用B超影像定位肝癌病灶区域有助于医生观察肿瘤及周围组织的特征变化,对病人的临床诊断、治疗方案选择以及预后具有重要的意义。然而在临床应用中,由于肿瘤的异质性和B超影像具有噪声斑点多、对比度差及病灶边缘不清晰等问题,导致定位结果高度依赖于放射科医生经验,具有较大的主观性。基于深度学习的方法建立了SSD肝癌病灶定位网络,并引入Yolo V3、Yolo V4、Faster RCNN和CenterNet等4种常见的目标检测网络进行对比实验,利用自主建立的肝癌B超影像数据集训练和评估模型性能。结果表明,相比其他网络,SSD各方面性能更好,能够自动、快速定位肝癌目标,准确率为94.8%,平均交并比为78.4%。Using B-ultrasound images to locate the focal area of liver cancer is helpful for doctors to observe the characteristic changes of tumor and surrounding tissues, which is of great significance for clinical diagnosis,treatment plan selection and prognosis of patients. However, in clinical application, due to tumor heterogeneity and the problems of B-ultrasound images such as multiple noise spots, poor contrast and unclear edges of lesions, the localization results are highly dependent on doctors’ experience and have great subjectivity. In this study, the SSD network was established for localization of liver cancer focus based on deep learning method, and four common target detection networks, Yolo V3, Yolo V4, Faster RCNN and CenterNet, were introduced for comparative experiments. The model performance was trained and evaluated by using the self-established B-ultrasound data set of liver cancer. The results show that SSD has better performance in all aspects compared with other networks, and can automatically and quickly locate liver cancer targets, with an accuracy of 94.8%and an average crossover ratio of 78.4%.

关 键 词：深度学习 肝癌病灶定位 目标检测 SSD网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]