摘要
目的:针对腔镜手术中气泡边缘模糊、数量形态浮动较大所造成的图像分割困难的问题,提出一种基于极坐标变换和深度学习的腔镜图像气泡分割方法。方法:首先构建包含笛卡尔坐标系下的U-Net网络和极坐标系下的U-Net网络的级联U-Net模型;其次,由级联U-Net模型中的笛卡尔系网络初步分割气泡,之后进行连通域分析并提取质心,然后以质心为原点将输入图像转换到极坐标系下;再次,由极坐标系网络分别预测气泡分割结果并对结果加权求和从而得到最终的预测结果;最后,为验证级联U-Net模型的有效性,对比级联U-Net模型与U-Net模型、深度轮廓感知网络(deep contour-aware network,DCAN)模型和边缘感知网络(edge-aware network,EAN)模型3种经典图像分割模型在腔镜气泡数据集上的分割效果。结果:级联U-Net模型对于腔镜图像气泡分割的精度显著优于3种经典图像分割模型,Dice系数、平均交并比、准确率、召回率分别为0.891、0.812、0.920和0.871。结论:基于极坐标变换和深度学习的腔镜图像气泡分割方法可精确分割腔镜图像中的气泡,可为腔镜手术中气泡自动去除装置的设计提供重要参考。
Objective To propose an endoscopic image bubble segmentation method based on polar coordinate transformation and deep learning to solve the problems due to ambiguous edges and large changes of numbers and shapes of bubbles during endoscopic surgeries.Methods A cascaded U-Net model was established involving in two U-Net networks in the Cartesian coordinate system and polar coordinate,respectively;the bubbles were preliminarily segmented in the Cartesian network,the connected components were analyzed to extract their centroids and then the centroids were used as the origins to convert the images to the polar coordinate;the bubble segmentation results were predicted separately using the polar coordinate system,and then were weighted and summed to obtain the final result;three classical image segmentation models including the U-Net model,deep contour-aware network(DCAN)model and edge-aware network(EAN)model were compared with the cascaded U-Net model in terms of the segmentation result for the endoscopic bubble dataset to verify its effectiveness.Results The cascaded U-Net model significantly outperformed the three classical image segmentation models when used for segmenting endoscopic image bubbles,with the Dice coefficient,average intersection and merger ratio,accuracy and recall rate being 0.891,0.812,0.920,and 0.871,respectively.Conclusion The endoscopic image bubble segmentation method based on polar coordinate transformation and deep learning behaves well and provides references to design bubble auto-elimination for endoscopic surgeries.
作者
周阳
顾伟
张杰
戴伟
李大永
胡洁
胡袁哲
ZHOU Yang;GU Wei;ZHANG Jie;DAI Wei;LI Da-yong;HU Jie;HU Yuan-zhe(Purchasing Center of Shanghai Chest Hospital,Shanghai Jiao Tong University School of Medicine,Shanghai 200030,China;Thoracic Surgery Department of Shanghai Chest Hospital,Shanghai Jiao Tong University School of Medicine,Shanghai 200030,China;Shanghai Jiao Tong University School of Mechanical Engineering,Shanghai 200240,China)
出处
《医疗卫生装备》
CAS
2023年第8期10-15,共6页
Chinese Medical Equipment Journal
关键词
极坐标变换
深度学习
腔镜图像
气泡分割
腔镜手术
polar coordinate transformation
deep learning
endoscopic image
bubble segmentation
endoscopic surgery
作者简介
周阳(1985-),男,硕士,工程师,主要从事医疗设备全生命周期管理方面的研究工作;通信作者:胡袁哲,E-mail:huyuanzhehyz@sjtu.edu.cn。
