Research On Liver Tumor Segmentation Algorithm Based On Deep Learning

As one of the most common sites for tumors in the human body,the liver is an important metabolic organ.Early screening,diagnosis,and treatment of liver tumors are crucial,as most liver tumors are malignant.Liver tumor segmentation on CT images can serve as a valuable basis for subsequent treatment.However,traditional liver tumor segmentation methods require manual labeling of patients’ CT images,which is a time-consuming and labor-intensive task.Therefore,computer-aided diagnosis is necessary to improve diagnosis efficiency.However,due to the complex distribution,various shapes,sizes,and numbers of liver tumors,automatic segmentation algorithms for liver tumors remain a challenging topic.In this paper,we provide a systematic overview of existing liver tumor segmentation algorithms based on CT images,analyzing their advantages and disadvantages,summarizing their shortcomings and challenges,and proposing a new liver tumor segmentation algorithm.The main research objectives are as follows:1.Data augmentation and data preprocessing.For the liver tumor segmentation sample set is less,the experimental sample set is expanded by using spatial geometric changes.And perform contrast limited adaptive histogram equalization processing on the experimental data set to enhance the contrast between the target area and other organs,suppress the irrelevant noise in the background,and realize the preprocessing of CT images.2.Liver tumor segmentation algorithm based on Attention-Res UNet.Aiming at the redundancy issue of the encoding and decoding structure of U-Net,as well as the problem of insufficient feature extraction.To overcome these challenges,Attention-Res UNet introduces identity mapping in the feature extraction module to enhance the transmission of features,ensuring the effectiveness of feature extraction.It also uses bottleneck layer design to reduce the number of model training parameters.Additionally,this method incorporates an attention mechanism into the residual block,which allows for the extracted features to focus on spatial information.By recalibrating the feature channels,the weight of effective features is strengthened,thus improving the segmentation performance.3.To further strengthen feature transfer,this paper designs Attention-Dense UNet to segment liver tumors.By designing the Attention-Dense Block to replace the convolutional layer in U-Net to strengthen feature propagation.The features in the Attention-Dense Block layer are integrated with the features of all feed-forward layers to obtain multi-level information,thereby enhancing the representation ability of the model.The design of dense blocks not only strengthens the reuse of features and improves the segmentation performance,but also greatly reduces the number of parameters of the model.Furthermore,by incorporating an attention mechanism within densely connected layers,the transfer of effective features can be enhanced to optimize segmentation results.At the same time,aiming at the class imbalance problem of the sample set,this paper proposes a hybrid loss function combining binary cross-entropy loss function and Tversky loss function to speed up the training of the model and further reduce the mis-segmentation rate.Attention Res U-Net and Attention Dense U-Net achieved competitive results in the3D-IRCADb and Li TS datasets compared to other segmentation algorithms,as measured by evaluation metrics such as the Dice coefficient.Ablation experiments confirmed the effectiveness of our proposed model structure and loss function.