Research On 3D Segmentation Of Liver And Tumor Based On Deep Learning

With the rapid development of deep learning and medical image processing technology,the application of deep learning models in the medical field has increased rapidly.CNN-based algorithm research has achieved remarkable results and is increasingly applied in the field of medical image analysis,improving the work efficiency of clinicians.In recent years,research on 3D CNN has received extensive attention.3D CNN makes up for the deficiency of 2D CNN’s insufficient ability to extract information in space,and can realize direct input and output of 3D medical data.This paper briefly introduces the development process of CNN from 2D to 3D,and uses the 3D CNN model to realize the segmentation of liver and liver tumors.Automated segmentation of liver tumors plays an important role in the diagnosis and treatment of liver cancer.Since the computed tomography images are mostly in3 D format,we design a fully 3D-based CNN model on the U-Net architecture for liver and liver tumor segmentation in two stages.In the first stage,liver segmentation is performed using a 3D U-Net model combined with residual structure,and the segmentation result is extracted as a region of interest(ROI).In the second stage,the morphological contour is cropped for the ROI area,and the improved 3D ResUnet network is used to accurately segment the tumor in this area.The attention mechanism,hole convolution and dynamic convolution ideas are introduced between the encoder and the decoder,which improves the feature extraction ability of the model for small targets.For data-driven deep learning models,the processing of the input data is more important than the design of the model.In this paper,a variety of data augmentation strategies are designed to solve the problems of blurred boundaries and different volumes of segmentation objects.The contrast between the liver area and the background area is increased by means of window width and window level adjustment and histogram equalization.The color flip is used to enhance the neural network’s attention to the tumor area.The ROI cropping is used to reduce the redundancy of background information.Ensuring that the direct input data of the neural network is compact and efficient.In addition,we use transfer learning in the training process to transfer the information learned in the liver segmentation task to the tumor segmentation task,which solves the problems of insufficient training data and class imbalance.Enhances the model’s performance while improving the model fitting speed and robustness.Verified by the LiTS dataset,the Dice coefficient of the liver segmentation task is 94.9%,and the Dice coefficient of the tumor segmentation task is 53.2%.Compared with the basic network framework,the tumor segmentation task can improve the segmentation performance by 4.4% on average.The experimental results show that the 3D CNN model designed in this paper successfully completes the liver and liver tumor segmentation tasks.At the same time,the effectiveness of modules such as attention mechanism,dynamic convolution and atrous convolution applied to3 D CNN model is proved.It provides a certain reference for the design and optimization of the 3D neural network model.