双通道Unet模型对前列腺自动勾画的研究

Study on the automatic sketch of DUnet model on prostate

目的:在图像分割的深度学习算法Unet网络基础上建立双通道Unet(DUnet)深度学习结构模型,以提高前列腺器官图像自动分割的准确性。方法:选取100例受检者的前列腺核磁扫描图像,其中50例来自医院图像资料系统,50例来自国际医学图像计算和计算机辅助干预(MICCAI)Grand Challenge数据库。100例受检者的扫描图像中81例为训练集,10例为验证集,9例为测试集。采用相干增强扩散(CED)算法对原始图像纹理和边缘进行强化,通过跳跃连接突出有效特征,获取更多多维信息增加上采样分辨率。建立双通道收缩路径和扩张路径形成对称结构DUnet,并行提取和学习原始图像以及CED图像特征,将双通道输出特征图融合得到分割图像。采用整体的Dice系数(Accuracy)、以扫描对象为单位的Dice相似系数平均值(Mean DSC)、Dice相似系数中位值(Median DSC)、平均表面距离(ASD)、最大对称表面距离(MSD)和相对体积差(RVD)6项指标对DUnet、Unet^(#)(原始图像)和Unet^(*)(CED图像)3种方法进行评估。结果:DUnet、Unet^(#)表现均优于Unet^(*)。表现最好的DUnet相较于Unet^(#),Accuracy提高1.28%,Mean DSC提高1.43%,Median DSC提高0.86%,ASD降低0.2 mm,RVD降低2.66%。直观勾画方面,DUnet自动勾画与医生手动勾画吻合度更高,勾画更加精准。DUnet在前列腺器官边界起伏区域更能捕捉到其形状的变换,对混淆性、相似性边界区域也有更好的辨别。结论:DUnet模型在突出其纹理和边缘强化特征的同时,弥补了强化效果导致精细结构的损失,在前列腺图像分割与勾画方面较Unet具有更优的表现。

Objective:To build structural model of dual-channel Unet(DUnet)deep learning on the basis of Unet network of deep learning algorithm with image segmentation,so as to improve the accuracy of automatic segmentation for the image of prostate tissue.Methods:The scan images of nuclear magnetism of prostate of 100 subjects were selected.Among of them,50 cases were selected from information system of the images of hospital,and 50 cases were selected from Grand Challenge database of medical image computing and computer assisted intervention(MICCAI).In the scan images of 100 subjects,81 cases were in training set,and 10 cases were in validation set,and 9 cases were testing set.The algorithm that textures and boundary of original images were intensified by coherent enhancement diffusion(CED)was adopted.The available features were reinforced by jumping connections so as to obtain more multidimensional information and increase the upsampling resolution.Dual-channel contraction path and expansion path were established to form symmetrical structure DUnet,and to extract and learn the original images and CED image feature in parallel.And the outputted feature images from dual-channel were merged to obtain segmentation images.The integral Dice coefficient(Accuracy),the mean of Dice similarity coefficient(DSC)that used scanned objects as unit,the median of DSC(Median DSC),the average surface distance(ASD),the maximum symmetrysurface distance(MSD)and relative volume difference(RVD)were adopted to assess the three methods included DUnet,Unet^(#)(original images)and Unet^(#)(CED images).Results:The appearances both DUnet and Unet^(#)were better than that of Unet^(*).Compared with that of Unet^(#),the Accuracy of DUnet which appearance was the best was increased 1.28%,and the Mean DSC of that was increased 1.43%,and Median DSC of that was increased 0.86%,and ASD of that was decreased 0.2 mm,and RVD of that was decreased 2.66%.At the aspect of direct sketch,the goodness of fit of DUnet automatic sketch and manual sketch of doctors were higher,and the sketch of that was more precision.In addition,the DUnet could better capture the transformation of shape in the undulating region of boundary of prostate tissue and could better distinguish the boundary region with confusion and similarity.Conclusion:The proposed DUnet model of this study not only highlights the reinforcement features of texture and boundary but also makes up the loss of fine structure caused by the strengthening effect,which has better performance than Unet in segmentation and sketch of prostate images.