Enhanced Liver Imaging Reporting and Data System (LI-RADS) Through Multi-task Convolutional Neural Networks

The liver imaging reporting and data system (LI-RADS) is a comprehensive system for standardizing the terminology, technique, interpretation, reporting, and data collection of liver imaging. This system aims to develop a standardized process for evaluating the malignancy of liver tumors and reduce variability in interpretation of physician. However, doctors have to spend a long time due to complicated procedures for making judgments based on LI-RADS. Recently, some studies have used deep learning to perform liver lesion classification using LI-RADS, ensuring the possibility of intelligent liver lesion grading. However, current methods are inadequate for clinical application due to their limited accuracy. Furthermore, these methods mainly provide classification results without transparent rationale, making them less reliable for physicians, hindering their adoption in practical medical settings. In this chapter, we present a reliable and intelligent LI-RADS system employing multi-task convolutional neural networks. Manual LI-RADS grading is mainly judged according to four specific pathological features, therefore, this study introduces a multi-task model designed with four distinct branches, each dedicated to the extraction and analysis of these pathological attributes. This structure allows for the generation of detailed and effective deep features, directly enhancing the task of LI-RADS grading and consequently improving the model’s overall performance. Moreover, each branch provides the analysis results of the major features. In practice, the results of major features can be provided to doctors accompanied by the LI-RADS grading results, providing a robust basis for validating the grading’s reliability as determined through intelligent analysis. This study is the first to propose a multi-task deep learning framework to achieve reliable and high-precision classification results for LI-RADS.