基于级联卷积神经网络的前列腺MR图像自动分割算法研究

Study on the automatic segmentation algorithm of MR image of prostate based on cascade convolutional neural networks

目的:提出一种基于语义分割(U-Net)网络和残差网络(Res-Net)两级神经网络级联的前列腺MR图像自动分割算法,以精确分割前列腺MR图像,准确评价前列腺癌诊断和疗效。方法:采用一级和二级卷积神经网络分割前列腺MR图像,一级网络采用二维U-Net网络初步定位和分割前列腺组织;二级网络联合使用U-Net网络和Res-Net网络精确分割前列腺组织。网络训练数据选自Github开源数据库中116例前列腺MR的T_(2)加权成像(T_(2)WI)数据集,其中80%用于训练,10%用于验证,10%用于测试。网络定量评价指标采用前列腺自动分割区域与手动分割区域之间的Dice相似性系数、灵敏度和特异度。结果:两级神经网络级联的前列腺MR图像自动分割算法在测试集所得的前列腺分割区域的平均Dice相似性系数为93.8%,灵敏度为94.6%,特异度为99.3%。结论:两级神经网络前列腺自动分割算法能显著提高前列腺MR的T_(2)WI图像分割精度,是一种具有较高稳健性和重复性的前列腺图像自动分割算法。

Objective:To propose an automatic segmentation algorithm of magnetic resonance(MR)for prostate based on cascading of two-level neural networks with semantic segmentation(U-Net)network and residual network,so as to precisely segment MR image of prostate and accurately evaluate the diagnosis and therapeutic effect for prostate cancer.Methods:The one-level and two-level convolutional neural networks were adopted to segment MR image of prostate.The one-level net adopted two-dimensional(2D)U-net to preliminary position and segment the prostate tissue.The two-level net jointly used U-Net and Res-Net to precisely segment the prostate tissue.The data of net training set selected from the data set of T_(2)weighted imaging(T_(2)WI)of the prostate MR of 116 cases in Github open source database.In these data,80%of them was used to train,and 10%of them was used to verify,and 10%of them was used to test.The Dice similarity coefficient(DSC),sensitivity and specificity between automatic segmentation area and manual segmentation area for prostate were used for quantitative evaluation of the network.Results:The average DSC value,sensitivity and specificity of automatic segmentation algorithm of cascade two-level neural network for prostate MR image in the segmentation area of prostate from test set were respectively 93.8%,94.6%and of 99.3%.Conclusion:The automatic segmentation algorithm of two-level neutral network for prostate can significantly improve the precision of segmenting T_(2)WI image of prostate MR,which is one kind of automatic segmentation algorithm with higher robustness and repeatability for prostate image.