基于三维卷积神经网络的肋骨CT图像的断骨分割  

Broken Bone Segmentation of Rib CT Image Based on Three-dimensional Convolutional Neural Network

作　　者：刘国华[1] 董芮 Liu Guohua;Dong Rui(Tianjin Key Laboratory of Optoelectronic Sensor and Sensing Network Technology,College of Electronic Information Technology and Optical Engineering,Nankai University,Tianjin 300350,China)

机构地区：[1]南开大学电子信息与光学工程学院,天津市光电传感器与传感网络技术重点实验室,天津300350

出　　处：《南开大学学报（自然科学版）》2023年第2期17-21,共5页Acta Scientiarum Naturalium Universitatis Nankaiensis

基　　金：中央高校基本科研业务费专项资金。

摘　　要：针对医学图像数据少、分割目标大小不一、相较背景占比低而带来的无法精准分割的问题,在传统U-Net网络基础上,提出3D RSPU-Net网络模型结构.使用残差网络代替神经元作为基本单元结构,减少网络模型训练中梯度消失退化的问题;U-Net上采样和下采样结构单元增加挤压和激发模块,提取重要特征,抑制无关特征;引入批量标准化层,网络损失函数空间更加平滑,有泛化能力,增加了模型的鲁棒性;采用三维卷积代替普遍使用的二维卷积,充分利用数据空间信息.在182张肋骨CT图像数据集上,使用该网络模型IOU为0.846,豪斯多夫距离为5.236 mm,精准度0.853.相较于使用传统U-Net网络和其他现有模型,该模型在肋骨骨折分割方面有更好的准确度和可行性.To solve the problem of inaccurate segmentation caused by less medical image data, different size of segmentation targets and the low proportion of background, a 3D RSPU-Net network model structure is proposed based on traditional U-Net network. By applying residual network as the basic unit structure instead of neuron, the problem of gradient disappearance and degradation in network model training is reduced;Extrusion and excitation modules are added in U-Net up-sampling and down-sampling structural units to extract important features and inhibit irrelevant features;The network loss function space becomes smoother with generalization ability by introducing batch standardization layer, which increases the robustness of the model;All two-dimensional convolution operations are replaced by three-dimensional convolutions. On 182 rib CT image datasets,the IOU, Hausdorff distance and accuracy of the network model are 0.846, 5.236 mm and 0.853, respectively.It has better accuracy and feasibility in rib fracture segmentation than traditional U-Net network and other existing models.

关 键 词：图像处理技术 医学图像分割 肋骨CT 三维卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]