Research On Liver Image Segmentation Based On Deep Convolutional Neural Network

The liver segmentation task is to segment the liver from liver CT data.Because of the low contrast,high noise and pathological abnormalities between liver and adjacent organs in CT images,it is difficult to segment liver.With the rapid development of visual correlation algorithm,deep convolution neural network has been more and more widely used in medical image segmentation.Among them,the U-Net model,which captures the contraction path of context and allows accurate localization of symmetrically extended paths,achieves better segmentation results.However,simple U-Net can't work good enough;and simply stacking convolution cores to improve the effect of the model will increase the amount of parameters and calculation,which makes training and reasoning difficult.In order to improve the performance of U-Net,using residual module to replace convolution module to accelerate the convergence speed of the model.Then,in order to restrain the imbalance of pixel categories,using inverse of DICE coefficients as loss function,and the method based on morphology is introduced to weigh the pixels.The model of 3D liver segmentation is improved.Finally,random affine transformation and random elastic deformation are used to enhance the data.The experimental results show that the method improves the segmentation effect of 2D image obviously and has strong scalability on 3D data.In order to reduce the parameters and computation of U-Net,this paper firstly uses kU-Net composed of two cascaded small networks to improve the performance of the model,while reducing the parameters of the model;secondly,it uses a roughto-fine segmentation strategy to improve kU-Net,and then gets W-Net,which improves the performance of the model remarkably while hardly improving the parameters and computation of the model.After that,the extension path of W-Net is simplified and the shrinkage path of the model is fine-tuned,which can significantly reduce the amount of calculation and parameters of the model,while maintaining the segmentation effect basically unchanged.The experimental results show that the improved method based on the residual module proposed in this paper improves the liver segmentation effect significantly.Using kU-Net and the coarse-to-fine segmentation strategy can reduce the amount of parameters and computation while ensuring the performance of segmentation,which has good practicability.