基于深度迁移学习的心脏MRI图像左心室分割  

Left Ventricular Segmentation in Cardiac MRI Images Based on Deep Transfer Learning

作　　者：苑金辉 乔艳 费烨琳 胡晓飞[2] YUAN Jin-hui;QIAO Yan;FEI Ye-lin;HU Xiao-fei(School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;School of Geography and Bioinformatics,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南京邮电大学地理与生物信息学院,江苏南京210003

出　　处：《计算机技术与发展》2021年第6期35-39,共5页Computer Technology and Development

基　　金：国家自然科学基金(61771251);江苏省社会发展重点项目(BE2016773)。

摘　　要：在研究基于深度学习的左心室分割方法时,需要足够的有标注的图像,才能获得准确的分割结果,而有标注的左心室图像往往难以获得。因此,提出了一种基于迁移学习和多尺度判别的生成对抗网络(TLMDB GAN)的MRI左心室图像分割方法,解决心室图像数据不足的问题。模型包含一个分割网络和一个判别网络。分割网络(TLBSN)使用全卷积神经网络,利用迁移学习逐层微调辅助分割,判别网络是一个多尺度的判别网络,监督生成网络更好地学习图像的特征信息。实验结果表明,基于多伦多市儿童病医院影像科提供的数据集对左心室内膜和外膜分割Dice相似系数分别为0.939 9和0.969 7。对比其他分割模型,该模型明显提高了分割精度。In the study of left ventricular segmentation based on deep learning, sufficient labeled images are needed to obtain accurate segmentation results, while labeled left ventricular images are often difficult to obtain. Therefore, a left ventricular MRI image segmentation method based on transfer learning and generation confrontation network is proposed to solve the problem of insufficient ventricular image data. The model consists of a segmentation network and a discrimination network. The segmentation network uses the full convolution neural network and the transfer learning to fine-tune the auxiliary segmentation layer by layer. The discriminant network is a multi-scale discriminant network, and the supervised generation network can learn the feature information of the image better. The experiment shows that the Dice similarity coefficients of left ventricular endocardium and epicardium segmentation are 0.939 9 and 0.969 7,respectively, based on the data set provided by the Imaging Department of Sick Children’s Hospital of Toronto. Compared with other models, the proposed model has significantly improved the segmentation accuracy.

关 键 词：迁移学习 生成对抗网络 心脏MRI 左心室分割 多尺度 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]