ARTIFICIAL INTELLIGENCE FOR FEATURE RECOGNITION AND LI-RADS CLASSIFICATION ON COMPUTED TOMOGRAPHY AND MAGNETIC RESONANCE IMAGING: A SYSTEMATIC REVIEW OF DIAGNOSTIC PERFORMANCE AND EXTERNAL VALIDATION

Abstract

Artificial intelligence may reduce variability and workload in computed tomography and magnetic resonance imaging assessment of liver observations, but high internal performance does not establish transportability or clinical benefit. This systematic review evaluates artificial-intelligence methods explicitly linked to Liver Imaging Reporting and Data System feature recognition, category assignment, or category-informed diagnosis, with emphasis on diagnostic performance and external validation. PubMed was searched through October 3, 2026 using prespecified LI-RADS, artificial-intelligence, and cross-sectional-imaging concepts. Seventy-three records were identified and eighteen original studies met the operational eligibility criteria. Evidence included radiomics, conventional machine learning, convolutional neural networks, transformer architectures, self-configuring segmentation systems, post hoc explainability methods, and a report-based large language model. Performance was not directly poolable because studies addressed different targets: arterial-phase hyperenhancement, washout, enhancing capsule, ancillary features, LR-3 versus LR-4/5 discrimination, multicategory grading, LR-M subtyping, lesion detection, or hepatocellular carcinoma diagnosis. Several internal studies reported high area under the receiver operating characteristic curve or classification accuracy, but the strongest direct category-classification pilot fell from 60.4% internal accuracy to 41.2% and 47.7% in external computed-tomography and magnetic-resonance-imaging cohorts. A stepwise magnetic-resonance-imaging system also demonstrated external degradation for individual major features, particularly washout and capsule, despite relatively stable category accuracy. More recent multicenter studies showed improved robustness and clinically meaningful reader assistance, but true site-independent validation, calibration, prospective impact analysis, subgroup fairness, and failure-mode reporting remained uncommon. Feature-guided systems were generally more interpretable than end-to-end category predictors, while report-based language models answered a different question from image-level recognition. The available evidence supports AI as an adjunct for standardized feature assessment and reader support, not autonomous replacement of LI-RADS interpretation. Clinical readiness requires independent multivendor validation, pathology or rigorous longitudinal reference standards appropriate to the target, locked algorithms, calibrated outputs, human–AI workflow evaluation, and transparent reporting of exclusions, indeterminate cases, and domain shift.