基于深度学习的肝囊型包虫病超声图像中小病灶检测方法研究  

Research on deep learning-based detection method for identifying small lesions in ultrasound images of hepatic cystic echinococcosis

作　　者：米吾尔依提·海拉提 热娜古丽·艾合麦提尼亚孜 卡迪力亚·库尔班 严传波[2] Miwueryiti·HAILATI;Renaguli·AIHEMAITINIYAZI;Kadiliya·KUERBAN;YAN Chuanbo(College of Public Health,Xinjiang Medical University,Urumqi 830011,China)

机构地区：[1]新疆医科大学公共卫生学院,乌鲁木齐市830011 [2]新疆医科大学医学工程技术学院

出　　处：《临床超声医学杂志》2024年第2期163-170,共8页Journal of Clinical Ultrasound in Medicine

基　　金：国家自然科学基金项目(81560294);省部共建中亚高发病成因与防治国家重点实验室开放课题(SKLHIDCA-2020-YG2);一流本科课程建设专项经费(010302010112)。

摘　　要：肝包虫病是一种呈全球性分布的人畜共患性疾病。超声作为该病的首选诊断方法,虽能及时发现大病灶位置并进行评价,但对早期小病灶的检测能力不佳。本文基于经预处理的高质量肝囊型包虫病超声图像小病灶数据集,提出了一种基于YOLOv7的检测肝囊型包虫病5类分型超声图像中小病灶的方法,以实现肝包虫病的自动检测,提高临床诊断效率。首先,用硬件感知神经网络EfficientRep替换原特征提取主干,实现在保证精度和速度不受影响的前提下,提高对硬件设备的适配度;其次,用更优的WIoU(Wise-IoU)替换CIoU(Complete Intersection over Union),改善了YOLOv7网络的评价指标CIoU在作为损失函数时,梯度计算效果差,导致检测精度下降的问题;最后,在主干的最后第4层加入CBAM注意力,进一步提高了模型检测精度。本文在自建的肝囊型包虫病超声图像小病灶数据集上进行了训练,结果显示,改进后的模型平均精度均值为88.1%,相较原始的模型性能得到了提升,并超过了对比的其余主流检测方法。说明本模型能更高效地检测并分类肝囊型包虫病超声图像中小病灶的位置和类别,应用于临床上能节约医师资源、缩短报告时长、提高诊断效率。Hepatic hydatid disease is a globally distributed zoonotic disease.The ultrasound,being the preferred diagnostic method for this disease,although capable of timely detection and evaluation of major lesion locations,exhibits limited efficacy in detecting early small lesions.Therefore,this study utilizes a preprocessed high-quality dataset of small lesions in ultrasound images of liver cystic hydatid disease and proposes an improved YOLOv7-based method for detecting small lesions in the 5 classification ultrasound images of liver cystic echinococcosis to achieve automatic detection and improve clinical diagnostic efficiency.Firstly,replace the original feature extraction backbone with EfficientRep,a hardware-aware neural network,to enhance adaptability to hardware devices without compromising accuracy and speed.Secondly,the replacement of CIoU with a more refined WIoU(Wise-IoU)can improve the evaluation metric CIoU(Complete Intersection over Union)of YOLOv7 Network.However,this modification adversely affected gradient calculation performance and led to a decline in detection accuracy when employed as a loss function.The CBAM attention module was finally incorporated into the fourth layer of the backbone,resulting in a further enhancement of model detection accuracy.This study was trained on a self-built dataset of small lesions in ultrasound images of hepatic cystic echinococcosis.The results demonstrate that the modified model achieved an improved detection accuracy of 88.1%in terms of mAP in comparison to the original model,thereby outperforming other mainstream detection methods.The proposed model demonstrates improve the efficacy in the detection and classification of lesions in ultrasound images of hepatic cystic echinococcosis,thereby facilitating its application in clinical settings.This innovative approach optimizes physician resources,reduces reporting time,and improves diagnostic efficiency.

关 键 词：肝囊型包虫病 深度学习 目标检测 YOLOv7 EfficientRep Wise-IOU CBAM 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] R532.32[自动化与计算机技术—控制科学与工程]