Segmentation Of Hip Joint With Prosthesis Based On Deep Learning

After the artificial hip has been in use for some time,it may need to be revised.Separating the non-prosthetic portion of the CT image containing the hip prosthesis and assigning different labels to different bone tissues is necessary for revision of hip surgery planning.Automated and accurate segmentation of clinically needed tissues will provide a powerful assistance to doctors.In view of the achievements of deep learning in image segmentation,deep convolutional neural network is used as the framework.On the choice of network model,in order to avoid possible network along with the depth deepening degradation situation,selected with residual mechanism of 3D-Res Unet network,and joined in the expansion of the network path segmentation module is used to merge network output characteristic figure,each layer to achieve the purpose of more finely divided hip bone.With metal prosthesis for data set of sample size is less,the amount of normal sample data without metal prosthesis more of the status quo of migration learning thought,first using the normal data without metal prosthesis to pre training model,then add the CT value metal prosthesis replacement training data,because if the data directly into the model to participate in the training,The network will not fit well because of the difference between the two types of data.In order to make the model fit the data with metal prosthesis well only through a small range of parameter adjustment,in the preprocessing process except metal CT value replacement;The mean and standard deviation of the metal-free prosthesis data were also used to standardize the data with metal prostheses,further narrowing the differences between the two types of samples.Dice similarity coefficient,sensitivity,specificity and Hausdorff distance were used to evaluate the segmentation results.Comparison experiments and ablation experiments were carried out with the other three models.The experiments show that the designed data preprocessing method,the addition of segmentation module and the transfer training method can improve the segmentation accuracy.