LI-RADS-aligned artificial intelligence for liver cancer diagnosis: methods, evidence, and clinical readiness

Liver tumor diagnosis relies heavily on imaging, and the liver imaging reporting and data system (LI-RADS) provides a structured framework for evaluating hepatocellular carcinoma (HCC) and related entities in at-risk populations. Artificial intelligence (AI) has rapidly expanded across CT, MRI, and ultrasound/CEUS, yet its clinical credibility depends on adherence to modality-appropriate tasks, robust validation, and alignment with LI-RADS semantics. This narrative review synthesizes methodological patterns, diagnostic performance, and readiness for clinical translation of AI systems designed for liver tumor characterization across major imaging modalities. We examine modality-task alignment-including CT-based differential diagnosis of HCC, intrahepatic cholangiocarcinoma, metastases, and benign mimics; LI-RADS feature detection and category assignment on MRI; LR-M disambiguation on MRI and CEUS; and surveillance-era triage on ultrasound. Evidence quality is appraised through external validation, reader studies, robustness analyses, calibration, and uncertainty reporting. A minimal reporting checklist is provided to support methodological transparency and facilitate reproducibility. Across modalities, AI systems show strong potential to enhance liver tumor diagnosis when they mirror radiologist reasoning, explicitly handle temporal enhancement dynamics, and incorporate clinically relevant priors. Translation into practice will require protocol-aware robustness, calibrated confidence estimates, size-stratified performance reporting, and broader multi-center validation. When developed under these principles, LI-RADS-aligned AI may meaningfully improve consistency, interpretability, and scalability in imaging-based liver cancer care.