Feature-Driven LI-RADS Classification in Liver Cancer: A CNN-Based Approach for CT Imaging Under Data Scarcity

Liver cancer is a major global health concern, driving interest in computer-aided diagnosis (CADx) to enhance liver tumor detection. This paper introduces a new approach for LI-RADS classification of liver cancer using multiphase CT imaging to improve diagnostic accuracy and interpretability for high-risk hepatocellular carcinoma (HCC) lesions. Although multiphase CT is essential for assessing focal liver lesions (FLLs), current AI-based LI-RADS methods often lack focused feature analysis and standardized interpretation. To address this, three slice selection strategies are proposed, targeting the most diagnostically relevant information based on lesion characteristics, thereby enhancing CNN model input quality. A multiphase classifier is also introduced, integrating key lesion features into a framework for LI-RADS score prediction, aligning with medical standards, and improving interpretability in clinical settings. Extensive validation on abdominal CT datasets shows this approach significantly outperforms existing methods in F1 scores and accuracy.