Segmentation Algorithm Of Liver Tumor Based On Deep Learning

Liver tumors are a common disease that seriously threatens human health,with poor prognosis and low survival rate.As an important part of liver cancer diagnosis and treatment,the accuracy of liver tumor segmentation will directly image the prognosis of patients.With the continuous development of computer computing power,computerassisted medicine is accepted by the majority of medical workers because of its accuracy and convenience.How to accurately segment liver tumors from CT slices has become a difficulty.Liver tumors vary in size,shape,location,and number,and the effects of CT contour blurring and noise interference are common in CT.Due to the complexity of liver tumor distribution,image processing algorithms based on traditional morphology and manual feature selection based on machine learning are not ideal for liver tumor segmentation.This paper proposes an end-to-end fully convolutional neural network.Compared with other semantic segmentation networks,this network has improved the segmentation index of CT modal liver tumors.Good image features determine the effectiveness of model training.First,data preprocessing and data enhancement are performed.In this paper,Li TS and 3Dirca Db CT modalities of abdominal three-dimensional CT scan data are selected and sliced,and slices with tumor distribution are selected for this article Training data and using random elastic transformation and random affine transformation for data enhancement,which makes the distribution of tumors in CT slices more complicated;later,in order to improve the network’s ability to segment tumors of different scales and tumor edges,on the benchmark network Attention-Unet A context information extraction structure is added,which consists of a DAC module based on hole convolution and a multi-core pooled RMP module.The DAC module ensures that the network can extract high-level semantic information while minimizing the loss of detailed information.The RMP module improves the network’s ability to extract multi-scale features,and to a certain extent,it combines detailed information and semantic information;next,Aiming at the problems that cross-entropy loss is not fine enough for tumor segmentation and Tversky loss convergence is slow and the training process is unstable,this paper proposes a dynamic weighted mixed loss function based on cross-entropy loss function and Tversky loss function,which can be passed through training The number of iterations is assigned to the weights of the two loss functions.In the early stage of training,the cross-entropy loss dominates the initial segmentation of liver tumors,and in the later period of training,the Tversky loss dominates the refinement of the tumor edges.In order to speed up the convergence of the model,we use the residual module and batch normalization to optimize the improved network results,and transfer learning by loading the pre-training weights of Res Net-34.After experimental testing,the network model based on the context information extraction structure proposed in this paper not only has a good segmentation effect on multi-scale liver tumors,but also can achieve fine segmentation of pixels on the tumor edge.In the currently published mainstream medical semantic segmentation network,the method proposed in this paper has been significantly improved in the segmentation evaluation indicators such as Dice coefficient.