Research On Segmentation And Detection Of Liver Tumors Based On Deep Learning

Liver cancer,a malignant tumor with a high fatality rate.As liver cancer is not obvious in the early stage,routine examination is difficult to detect,in the late recurrence rate is high and difficult to cure,so early detection is extremely important.Computed tomography(CT)and magnetic resonance imaging(MRI)can provide useful information for the diagnosis of liver cancer,but doctors need to review a wide range of medical images,which is time-consuming and can cause misdiagnosis due to fatigue.With the rapid development of deep learning algorithm,medical imaging technology combined with CT and deep learning is gradually applied to medical diagnosis.Through the algorithm model,liver and liver tumors can be automatically labeled to assist doctors to detect liver cancer as early as possible.Considering that in the early stage of liver cancer,the liver tumor size is small,the lesion boundary is blurred,and the tumor density is close to that of normal liver,etc.,this paper carried out research on liver and liver tumor recognition based on deep learning.The main research contents are as follows:(1)Medical data sets are established.In this paper,274 cases of liver cancer patients with abdominal CT images obtained from the imaging department of cooperative hospitals from July 2017 to January 2022 were retrospectively collected,and based on this,the cooperative hospital data set was constructed.Among them,the gold standard of data was marked by two imaging experts according to the clinical report,and the tumor location was also framed on the 18,460 abdominal pictures containing tumors.The above data labeling was checked by professional doctors twice to ensure that the labeling was correct.(2)A liver segmentation algorithm based on multi-scale attention is proposed.The model is improved on the basis of UNet: the convolutional blocks in the network are replaced with residual blocks to ensure that the performance of the deep network is not worse than that of the shallow network.Attention gating module is added at the jump joint to enhance the model’s focus on the target region.Use Drop Block regular blocks to enhance model robustness;With the help of multi-scale modules,the network has good segmentation performance on different sizes of liver.In addition,a hybrid loss function is designed,which not only alleviates the problem of class imbalance in the image,but also improves the convergence performance of the network.The experimental results showed that the above improvements could effectively improve the segmentation effect of liver,and the Dice coefficient was increased from 0.9259 to 0.9542.(3)An attention-based liver tumor detection algorithm is proposed.The model is improved on the basis of YOLOv5: a branch is introduced into the original SPP module to improve the detection effect of small tumors;Add CA attention module before SPP module to further improve the performance of the model;With the help of segmentation model and segmentation graph obtained from post-processing operation,the prediction frame is further screened.The experimental results show that the above improvements can effectively improve the detection effect of liver tumors,and the average accuracy rate(m AP)is increased from 0.916 to 0.928.