Method For Extracting Midsagittal Plane In Brain Based On Deep Learning

In the clinical diagnosis process,it is very important to extract the sagittal plane of the brain.Calculating the declination of the head through the mid-sagittal plane,thereby correcting the position of the head in the image,can provide the best viewing angle for diagnosing brain diseases;Calculating the symmetry of the left and right hemispheres of the brain based on the mid-sagittal plane can also provide an important reference for judging certain brain diseases.However,the existing mid-sagittal plane extraction algorithm usually has problems such as poor extraction effect,large calculation cost and poor robustness.Therefore,this topic proposes a mid-sagittal plane extraction scheme based on deep learning.The scheme is divided into three steps:(1)use the image segmentation network to identify key point areas;(2)use the k-means method to calculate each cut plane The key point coordinates in the middle,because the number of key points in each slice is 2,so k is set to 2;(3)According to the key point coordinates,a random sampling consistency algorithm is used to fit the middle sagittal plane,which can effectively Eliminate the interference of abnormal points.Based on U-Net,a new three-dimensional image segmentation network is proposed.The new network Group＿Unet has the following four characteristics:(1)Each channel in the three-dimensional CT image is treated as a two-dimensional image to form a channel,forming 24 channels of two-dimensional data input Group＿Unet;(2)Modify the channel splicing of U-Net to add channels to form Group＿Unet,so that it better reflects the one-to-one correspondence between end-to-end;(3)Use group convolution to balance the relationship information between channels And the geometric information inside the channel;(4)Modify the number of channels in each stage of Group＿Unet to adapt to the subject data.The experiment used 240 three-dimensional brain CT images marked by ourselves,including 199 samples in the training set and 41 samples in the verification set.The model is evaluated from three angles: image segmentation,key point detection and plane fitting.The 41 samples of the validation set are predicted.The Dice similarity coefficient of image segmentation is 0.768,the accuracy rate is 0.816,and the recall rate is 0.725.The normal error NE of the key point detection is 1.558,the deviation angle of the mid-sagittal plane of the plane fitting is 1.080 °,and the running time of the entire process is 0.534 seconds.The experiment proves that the newly proposed mid-sagittal plane calculation method has high accuracy,strong robustness and fast running speed,and can be well used in the diagnosis of brain diseases.