Liver Lesion Segmentation Based On Deep Learning

As one of the most critical organs in human body,liver undertakes the function of detoxification metabolism for the body,and the health of the liver is of great importance.China has always been a major country with a high occurence of liver cancer,and the incidence of liver cancer remains high.In addition,the survival rate of liver cancer is much lower in the late stages than in the early stages.Early detection of liver cancer can greatly improve the survival rate of patients,so early screening for liver cancer is quite imperative.Computed Tomography is currently the most common and effective medical imaging technology used in the detection of liver diseases.In the traditional way,early screening and preoperative diagnosis of liver cancer require doctors to understand the shape,size and location of the tumor with the help of CT,so as to complete the task of tumor segmentation.This takes time and effort and depends on the clinical experience of the doctor.Based on the above background,in order to efficiently and automatically complete the segmentation of liver lesions.In this paper,a liver lesion segmentation model based on deep learning is proposed and studied in depth.The main work and innovations of this this are presented as follows:(1)Construct a high-quality liver CT dataset.Current open datasets of liver CT are relatively scarce,while the existing datasets have some quality problems.For this reason,we cooperated with experts in the imaging department of third-class hospital to mark and obtained 200 sets of mixed abdominal CT images of plain scan and enhanced phase.Verified by various methods,the Dice score in the self-built dataset was 15% higher on average than the current best public liver CT dataset,reflecting the quality of the self-built dataset.(2)An improved U-net network structure is proposed to combine with Res Ne St network to improve the feature extraction capability of model.In addition,A domain adaptability module is proposed for self-built multi-phase mixed data.At the same time,a new attention mechanism lcs SE module is proposed,which can extract important features more accurately.The proposed network achieved a Dice score of 93.76%,an improvement of 0.11% over the current state-of-the-art approach.(3)In order to further improve the segmentation accuracy of liver lesions,a classification auxiliary task network structure and false positive filtering algorithm were proposed based on the above network and combined with multi-task learning,which greatly reduced the false positive cases in the model prediction and finally reached the Dice score of 94.23%.The above methods have been verified by a large number of comparison experiments and ablation experiments,and the results fully prove the validity and correctness of the proposed model and module.