基于深度学习的CT图像肋骨自动分割与三维重组研究

Research on Automatic Rib Segmentation and 3D Reconstruction of CT Image Based on Deep Learning Methods

目的探讨基于卷积神经网络的深度学习模型在胸部CT图像上对肋骨区域的自动分割与三维重组的价值。方法搜集2020年11月至2021年1月在本院行胸部CT检查者共130例(共计33280张轴位图像),以其中的80例作为训练集,20例作为测试集,来自另外三台不同CT设备的被检者各10例作为独立验证集,评价基于四种3D分割网络(UNet 3D、VNet、DenseNet 3D与DenseVoxelNet)的深度学习模型对CT图像上肋骨区域的分割效果,并将结果应用于肋骨三维重组。结果基于四种3D分割网络的深度学习模型在测试集上的最高Dice系数分别为99.91%、99.80%、99.90%、99.73%,准确率(IOU)分别达到99.82%、99.60%、99.80%、99.47%,其中UNet 3D网络在验证集上达到最高99.85%的Dice系数,IOU达到99.70%。结论基于卷积神经网络的深度学习模型在胸部CT图像上对肋骨区域的自动分割效果良好,并具有一定的泛化能力,分割结果应用于肋骨三维重组具有较高的临床价值。

Objective To explore the value of the deep learning model based on convolutional neural network in the au-tomatic segmentation ofthe rib region in chest CT image,discuss the feasibility of applying the results to 3 D reconstruction.Methods 130 chest CT examinations(33280 axial images in total) were collected from November 2020 to January 2021 in our hospital. 80 of them were used as training dataset,20 examineeswere taken as test dataset,30 examinees from other 3 different CT devices(NeuViz128、GE LightSpeed VCT、GE Optima 680)were taken as independent verification datasets,andevaluated the segmentation effect of the deep learning model based on 4 kinds of 3 D segmentation networks( UNet 3 D、VNet、DenseNet 3 D and DenseVoxelNet) on the rib area on the CT image,and apply the results to the 3 D reconstruction ofthe ribs.Results The highest Dice coefficients of the four kinds of 3 D segmentation networks were 99. 91%,99. 80%,99. 90%,99. 73% respectively. The IOU reached 99. 82%,99. 60%,99. 80%,99. 47% respectively. Among them,theUNet 3 D network reached the highest Dice coefficient of 99. 85% on the verification dataset,and the IOU reached 99. 70%.Conclusion The deep learning model based on the convolutional neural network has a good effect on the automatic seg-mentation of the rib area on the chest CT image,and has a certain generalization ability. The segmentation result hasa highclinical value when applied to the 3 D reconstruction of the ribs.