基于BiSeNet的小儿超声心动图左心分割方法

Pediatric Left Cardiac Echocardiography Segmentation via BiSeNet

小儿超声心动图分割是后续生物学参数测量与疾病诊断的关键一步。目前,这主要依赖于超声医生的手动分割,不仅耗时耗力,而且由于它的重复性与冗余性,常常会导致不准确的分割。深度学习方法在自然图像处理领域已经取得令人瞩目的成果,因此提出应用深度卷积神经网络,从小儿超声心动图中学习有效特征,进行左心关键解剖结构的分割。具体来说,提出使用双路径分割网络(BiSeNet),通过两路分支网络,分别提取低层和高层的特征,然后送入一个特征融合模块,筛选出有效的特征,从而得到准确的分割结果。在采集自深圳儿童医院超声科的包含87个超声心动图视频(2 216张图像)的数据集上进行验证,并与医生的标注结果进行比较。实验结果表明,BiSeNet可以提取到超声心动图中心脏结构的特征,它在左室和左房的分割任务上取得Dice系数高达0.914和0.887。这证明,所提出的方法可以帮助医生进行超声心动图分割,从而减轻医生的负担。

Accurate segmentation of pediatric echocardiography is an essential step for the later biomedical measurement and diagnose.Currently,it relies on sonographer′s manual segmentation,which is time consuming and redundant,and therefore may lead to mistakes.Deep learning methods have achieved remarkable results in the field of computer vision.Therefore,we proposed to extract features from pediatric echocardiography images via deep convolutional neural networks and segment key anatomical structures of the heart.Specifically,we used BiSeNet consisting of two components,spatial path and context path,to extract low and high level features,respectively,and then fused them via a feature fusion module to get the most important features,for accurate segmentation.We conducted experiments on a dataset consisting of 87 echocardiography videos(2216 images)collected from Shenzhen Children Hospital,and compared our prediction with sonographers′ground truth.Results showed that BiSeNet was able to capture the structure feature of echocardiography images,and achieved 0.914 and 0.887 in term of Dice index in left ventricle and left atrium segmentation task,respectively.The proposed method could help with accurate pediatric echocardiography segmentation,and released sonographers from redundant work.