Research On Segmentation Of Liver Tumors Based On Multi-scale Attention And Edge Enhancement

Liver cancer has become one of the most common diseases in humans,causing a large number of deaths every year.Segmentation and quantitative analysis of liver tumors are necessary steps for classification and treatment of liver cancer.Liver tumors are usually segmented manually by radiologists,which is not only time-consuming but also depends on the radiologist’s expertise.Therefore,the automatic segmentation algorithm of liver tumors is of great significance to assist doctors in clinical diagnosis and treatment.Compared with natural image segmentation,medical image segmentation is an extremely challenging task due to its inherent difficulties.Because the contrast between organs in medical images is relatively low,and the size,shape,and location of lesion areas within the same patient are all different.In addition,some tumors have fuzzy boundaries,which pose challenges for accurate segmentation tasks.Based on this,this paper has carried out key research on the automatic segmentation algorithm of liver tumors based on deep learning methods.The work content is as follows:(1)In this paper,some existing liver tumor segmentation algorithms are briefly reviewed,and the main research difficulties of liver tumor segmentation are discussed.The multi-scale problem and fuzzy boundary problem of liver tumors are the main difficulties in the development of liver tumor segmentation.Secondly,this paper proposes an automatic liver tumor segmentation method for these two research difficulties and has done a large number of ablation experiments and comparative experiments on the dataset related to the liver tumor segmentation to prove the effectiveness and superiority of the proposed model.Finally,in order to verify the generalization performance of the model in this paper,this paper tested the model on the clinical dataset provided by the Second Hospital of Shandong University,and the segmentation effect diagram confirmed that the model has good generalization.(2)This paper proposes a feature fusion residual network based method for liver tumor segmentation using deep supervision.The algorithms proposed in this paper are all fully convolutional neural network frameworks,using the "encoder-decoder" structure for segmentation.In order to fully decode the deeper semantic information of the data,a feature fusion module based on three deep features with different resolutions is proposed in the decoder part.At the same time,in order to make full use of all the hierarchical features of the encoder,including the detailed information of shallow features,this paper designs a deep supervision module for auxiliary supervision.Experiments have proved that this algorithm has a significant improvement in the Dice index compared with other fully convolutional neural network algorithms.(3)This paper proposes a liver tumor segmentation method based on a multi-scale attention mechanism and an edge enhancement module.In this paper,a hidden layer feature multi-scale attention fusion strategy based on convolutional channel attention is used to address the multiscale problem of tumors.The attention mechanism is added to the encoder path,so that the feature extraction pays more attention to the multi-scale information of the image,integrates the pyramid structure,and obtains the optimized feature information.In addition to the multi-scale problem,another major difficulty in liver tumor segmentation is that the boundaries of tumors are blurred.Low contrast and fuzzy borders lead to a large difference between the edge contours of tumor segmentation results obtained by many segmentation models and the actual values.To optimize this problem,this paper proposes an edge enhancement strategy guided by hidden layer feature confidence for encoder segmentation prediction.After the initial segmentation map is obtained,the confidence map is calculated for the initial segmentation probability map,and the accuracy of the edge contour is improved by continuously expanding the area of the determined pixels.Experimental data show that the algorithm has obvious advantages in medical segmentation indicators compared with other network models.In order to solve the multi-scale problem and fuzzy boundary problem of the liver tumor segmentation task,this paper proposes a liver tumor segmentation algorithm based on feature fusion residual network using deep supervision and a liver tumor segmentation algorithm based on multi-scale attention mechanism and edge enhancement.The superiority of the algorithm proposed in this paper is proved by the quantitative analysis of the segmentation index and the qualitative analysis of the segmentation result graph on the liver tumor segmentation related dataset.The test result graph on the clinical data set provided by Second Hospital of Shandong University proves the good generalization of the algorithm proposed in this paper.