第22卷第2期2023年2月Vol.22No.2Feb.2023软件导刊SoftwareGuidePRA-UNet3+:全尺度跳跃连接CT肝脏图像分割模型钟经纬(江南大学人工智能与计算机学院,江苏无锡214122)摘要:器官损伤死亡率高,严重威胁着人类的生命安全。人体内脏形态多样,解剖结构复杂,因此器官图像的准确分割有助于医生进行诊断。医学图像对高精度分割模型的需求很大,然而,大多数医学图像分割模型都是直接从一般的图像分割模型迁移过来的,常常忽略了浅层特征信息以及边界的重要性。为解决该问题,提出使用注意力门和点采样方法获得高质量分割边界的图像分割模型。在常用的肝脏医学图像数据集CHAOS上对该模型进行评估,平均Dice达到0.9467,平均IoU达到0.9623,平均F1Score达到0.9351,证明该模型可同时学习图像细节特征和全局结构特征,能更好地对肝脏图像进行分割。关键词:医学图像分割;U-Net;注意力门;点采样技术DOI:10.11907/rjdk.221218开放科学(资源服务)标识码(OSID):中图分类号:TP391.41文献标识码:A文章编号:1672-7800(2023)002-0015-06PRA-UNet3+:Full-scaleConnectedCTLiverImageSegmentationModelZHONGJing-wei(SchoolofArtificialIntelligenceandComputerScience,JiangnanUniversity,Wuxi214122,China)Abstract:Organlesionshaveahighmortalityrateandseriouslythreatenthesafetyofhumanlife.Theinternalorgansofhumanbodyaredi⁃verseinformandcomplexinanatomicalstructure,accuratesegmentationoftheorganassiststhedoctorinmakingthediagnosis.Highpreci⁃sionsegmentationmodelisrequiredformedicalimage.However,mostsegmentationmodelsaredirectlytransferredfromthegeneralimageseg⁃mentationmodel.Thesemodelsoftenignoretheimportanceofshallowfeatureinformationandboundaries.Inordertosolvethisproblem,atten⁃tionmechanismandpointsamplingtechniqueareproposedtoobtainhighqualitysegmentationboundary.ThemodelwasevaluatedonCHA⁃OS,acommonlyusedlivermedicalimagedataset,andtheaverageDicewas0.9467,theaverageIoUwas0.9623,andtheaverageF1Scorewas0.9351.Itisprovedthatthismodelcanlearnboththedetailfeaturesandtheglobalstructurefeaturesoftheimage,andcanperformbettersegmentationoftheliverimage.KeyWords:medicalimagesegmentation;U-Net;attentiongate;pointsamplingtechnique0引言近年...