Research On Medical Image Segmentation Technology Based On 3D Residual U-Net

In recent years,due to the deterioration of the ecological environment and irregular diet,the incidence rate of liver diseases and liver vessel diseases is increasing.In order to assist doctors in preoperative surgical planning,liver vessels segmentation is becoming more and more necessary.However,due to the low contrast of vessels region in CT images and the large differences in vascular geometry among different patients,it is very difficult to segment liver vessels accurately.With the increase of data volume and the improvement of computer computing capacity,deep learning methods based on convolutional neural network are favored by many scholars.Compared with traditional methods,deep learning can achieve end-to-end automatic segmentation.For medical image segmentation,FCN network and U-Net network are the most popular.In addition,in the segmentation of liver vessels,there are the following difficulties: 1)there are many small areas of liver vessels in the CT image,and in the process of sampling under the network,the characteristics of these vessel areas are easy to disappear,resulting in poor segmentation effect of small vessel areas;2)there is pixel imbalance between hepatic veins and portal veins in liver vessels.In the process of learning,the network pays more attention to the class with more pixels,resulting in poor segmentation effect of the class with fewer pixels.In view of the above problems,the main contents of this thesis are as follows:(1)The two-dimensional FCN and U-Net networks are used to segment liver vessels to observe the segmentation effect of two-dimensional neural network on liver vessels,and the segmentation results are quantitatively analyzed.(2)On the basis of two-dimensional neural network,a three-dimensional U-Net network is established,and to solve the problem of small vessel segmentation in CT images,the residual block in residual network is fused with 3D U-Net,and the segmentation performance of the network after adding residual block in different positions was explored.Finally,a 3D residual U-Net network with the best effect is established.Through the improvement of the network,the segmentation effect of liver vessels is improved significantly.(3)A weighted Dice loss function is proposed to solve the pixel imbalance between hepatic veins and portal veins.In addition,the effects of weighted Dice loss function,weighted cross entropy loss function and Focal loss function on pixel imbalance are compared.Finally,the experiment shows that the weighted Dice loss function has the best effect on alleviating the pixel imbalance.(4)The generalization and accuracy of the proposed method(3D residual U-Net network and weighted Dice loss function)are verified.Experimental results show that the model presented in this thesis has the best performance.The Dice coefficients of the proposed method in hepatic veins,portal veins and veins system are 71.7%,76.5% and 75.4%,respectively.Compared with the original 3D U-Net network and Dice loss function,the segmentation coefficients are increased by5.3,2.6 and 2.6 percentage points,respectively.In addition,compared with 2D FCN,2D U-Net network and other segmentation methods,the proposed method has the highest accuracy,which verifies the superiority of our proposed method.