Research Of Deep Learning Based CT Image Liver Tumor Segmentation Method

Liver cancer has the highest mortality rate among cancers,which is a serious threat to human health.The CT is one of the common ways for doctors to diagnose liver cancer in clinical applications,and the accurate segmentation of liver and tumor areas from CT images has important research value for the diagnosis and prognosis treatment of the disease.The results of traditional liver tumor segmentation algorithms mostly rely on the priori knowledge,which is not conducive to automated scaling in clinical applications.Deep Convolutional NeuralNetwork have been widely applied to liver tumor segmentation tasks in recent years,which allows the models to learn target features efficiently from a large number of samples through local perception and parameter sharing features.The task of liver tumor segmentation from CT images is still challenging due to the blurred edges and low contrast in the liver and tumor regions in CT images.The paper investigates the application of algorithms based on deep convolutional neural networks on task of liver and tumor segmentation from CT images,the main works as follows:1)Dedicated to the improvement of DCNN in terms of segmentation accuracy,Strip Pooling-Attention and Fusion Block is proposed.The model improves U-Net,which replaces the conventional convolution of the coding area with stripe pooling module and expands the local perceptual field of the network by the design of stacking flat-like convolutions.In the jump connection layer of SP-AFBNet,the mechanisms of multi-scale feature fusion and channel attention are added to overcome the problem of unbalanced feature information between different scales and strengthen the correlation between feature channels.The final results show that our algorithm can achieve better segmentation results on task of liver and tumor segmentation from CT images compared with other advanced algorithms.2)For the problems of high computational complexity and high inference delay of liver tumor segmentation model,the lightweight network LW-Mnet is proposed.The modified MobileNetv3 is used as the backbone network of LW-Mnet,which consists of a deep separable convolution and a non-local module,the former is the core of the model lightweight,and the latter captures global context-dependent information by calculating the similarity between any two units.To enhance the model multi-scale inter-feature information,the Atrous Spatial Pyramid Pooling is improved in terms of lightweight,i.e.,LW-ASPP.The decoding part of the model consists of the ShuffleNetv1 basic module,which extracts the main feature information and recovers the feature map resolution.The experimental results demonstrate that LW-Mnet outperforms other comparative algorithms both in terms of segmentation accuracy and lightweight.3)In order to implement the research results,an intelligent auxiliary diagnosis platform for liver tumor segmentation is designed and developed based on the above work.The frontend and back-end of platform are based on Vue and Django frameworks respectively,and the platform is mainly for clinicians and patients.The patients can use the platform to submit consultation forms and upload their CT images of liver parts,and doctors can complete the diagnosis of the diseased forms with the liver tumor segmentation function of platform.The construction of liver cancer consultation platform can effectively reduce the consultation cost of patients and improve the diagnosis efficiency of doctors.