LI-RADS Based Study On Grading Of Liver Lesions

In the recent century,the medical imaging technology has been gradually developed.The diagnostic process based on medical images has gradually become a basic diagnostic process.Recently,Radiological Society of North America began to develop standardized data reporting systems(RADS)for various cancer diseases,such as BI-RADS and LI-RADS.LI-RADS has been developed to standardize liver imaging findings and reports.It has been adopted in many medical organizations around the world.Although LI-RADS is of great significance,due to its complicated process,there is still no complete computer-aided diagnosis system based on LI-RADS.Due to the great significance of LI-RADS in the diagnosis of liver tumors and the important role of computer-aided system in medical image diagnosis,this paper studies and classifies the grading of liver tumors based on LI-RADS.For the complete LI-RADS,this paper gives a complete design of computer CAD flow.The designed CAD processing is divided into three main processes,including data preprocessing,classification of benign and malignant lesions and grade of malignancy.In data preprocessing,we design and implement a method that includes three steps:global registration,segmentation,and local registration.For the LI-RADS-based classification of benign and malignant tumors,the related work of LI-RADS defines a lot of characteristics supporting separation of benign and malignant lesions.However,these features are descriptive.Therefore,this paper quantifies these descriptive features based on LI-RADS so that the computer can automatically extract features from images.In order to design and extract more effective features,this paper summarizes the descriptive features in the diagnostic reports of radiologists from our database leading to a feature set based on both diagnostic reports and LI-RADS.The experimental results show that the combined feature set can solve the LI-RADS classification problem better.After the feature selection based on the greedy strategy and comparison of different traditional machine learning models,the AUC was 0.9483 based on the data set marked by radiologists.For the grade of malignancy,the LI-RADS standard defines the main features and the classification process.As the classification process is relatively fixed,the key of this problem lies in the quantification of the main features of LI-RADS.At present,the related quantification algorithms are not comprehensive in their data set.Their data set is lack of negative samples of one main feature.And also their algorithms are designed based on single phase.This paper first collects data based on every main feature,including positive and negative samples for all main features.Based on the data set,the quantification of the LI-RADS main feature based on multi-phase improves the AUC value by 7%compared with that based on single phase.Also a multi-scale strategy is adopted for the capsule which is the most difficult main feature to quantify and the AUC reaches 0.9013.