基于多种U-net模型的宫颈癌肿瘤超声影像靶区自动分割比较研究

Automatic segmentation comparison among multiple U-net models in the target delineation for cervical cancer based on ultrasound images

目的超声已被公认为是宫颈癌诊疗最常用成像方式之一,其准确分割是临床诊疗分析的基础,也是放射组学基本步骤。本研究比较不同U-net自动分割模型在基于超声影像宫颈癌肿瘤靶区的分割效果。方法采用U-net对基于超声影像的宫颈癌肿瘤靶区进行自动分割,以Resnet 34替换骨干网络方式提出一种逐层还原的新方法,并比较U-net、U-net++和U-net++基础上修改的多种模型分割结果。选取2017-12-01-2018-12-31温州医科大学附属第一医院收治的270例宫颈癌患者的544张超声图片,随机分成训练集408张、验证集54张和测试集82张,手动勾画由高年资超声诊断医生完成。用Dice相似性系数(Dice similarity coefficient,DSC)、Jaccard相似性系数(Jaccard similarity coefficient,JSC)和平均表面距离(average surface distance,ASD)对自动分割结果进行比较分析。结果各种U-net模型的自动分割结果与医生手动勾画结果相似度都比较高。各模型平均DSC、JSC和ASD分别为0.84~0.84、0.75~0.82和7.00~9.99。Resnet 34作为骨干网络的U-net使得原始U-net结果平均DSC从0.86提升至0.89,平均ASD从8.17降低至7.00,平均JSC从0.77提升至0.81,Resnet 34作为骨干网络的U-net获得了比原始U-net更好的分割效果。结论本研究提出的方法有助于提高U-net在超声图像的分割效果,U-net++with Resnet 34取得了最佳分割结果。

OBJECTIVE Ultrasound is the most frequently used imaging modality for the diagnosis and treatment of cervical cancer.Its segmentation is a critical step for cancer diagnosis,as well as for radiomics studies.The purpose of this study was to compare multiple U-net models in the automatic segmentation of targets for cervical cancer based on ultrasound images.METHODS A new layer-by-layer restoration method by replacing the backbone network with Resnet 34 was proposed in this study based on U-net++to automatically segment the targets of cervical cancer on ultrasound images.A total of 544 ultrasound images from 270 patients were collected and randomly divided into training set(408),validation set(54),and testing set(84).Manual delineations were conducted by a senior radiologist specialized in ultrasound imaging.Dice similarity coefficient(DSC),Jaccard similarity coefficient(JSC),and Average Surface Distance(ASD)were applied to evaluate the results of automatic segmentations.RESULTS High similarity was achieved for all these U-net models,although non-optical edge was observed for some algorithms.The average DSC,JSC and ASD for these models were around 0.84-0.84,0.75-0.82 and 7.00-9.99,respectively.U-net with Resnet 34 backbone improved the DSC and JSC from 0.86 to 0.89,0.77 to 0.81,and reduced the ASD from 8.17 to 7.00,respectively,compared with original U-net.CONCLUSIONS The suggested methods are able to improve the segmentation effects of U-net based automatic segmentation.The U-net++with the backbone of Resnet 34 achieved the best segmentation accuracy in the automatic segmentation task for cervical cancer patients based on ultrasound images.