A systematic review published in Lumen et Virtus in 2026 assessed artificial-intelligence methods for recognizing Liver Imaging Reporting and Data System (LI-RADS) features and assigning categories on computed tomography and magnetic resonance imaging. The authors searched PubMed through October 3, 2026, identified 73 records and included 18 original studies under prespecified eligibility criteria. The review examined diagnostic performance and external validation, the test of whether a system works beyond the data used to develop it.

The studies used radiomics, conventional machine learning, convolutional neural networks, transformer architectures, self-configuring segmentation systems, post hoc explainability methods and a report-based large language model. Their results could not be pooled because they addressed different tasks, including arterial-phase hyperenhancement, washout, enhancing capsule, ancillary features, LR-3 versus LR-4/5 discrimination, multicategory grading, LR-M subtyping, lesion detection and hepatocellular carcinoma diagnosis. Several studies reported high internal area under the receiver operating characteristic curve or classification accuracy, but the strongest direct category-classification pilot dropped from 60.4% internal accuracy to 41.2% in an external CT cohort and 47.7% in an external MRI cohort.

A stepwise MRI system also performed worse externally on individual major features, particularly washout and capsule, although its category accuracy was relatively stable. More recent multicenter studies reported greater robustness and clinically meaningful reader assistance. However, 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 addressed a different task from image-level recognition.

The review concludes that available evidence supports AI as an aid to standardized feature assessment and reader support, not as a replacement for LI-RADS interpretation. The authors say clinical readiness requires independent multivendor validation, appropriate pathology or rigorous longitudinal reference standards, locked algorithms, calibrated outputs, evaluation of human–AI workflows, and transparent reporting of exclusions, indeterminate cases and domain shift.

A systematic review published in Lumen et Virtus in 2026 assessed artificial-intelligence methods for recognizing Liver Imaging Reporting and Data System (LI-RADS) features and assigning categories on computed tomography and magnetic resonance imaging. The authors searched PubMed through October 3, 2026, identified 73 records and included 18 original studies under prespecified eligibility criteria. The review examined diagnostic performance and external validation, the test of whether a system works beyond the data used to develop it.

The studies used radiomics, conventional machine learning, convolutional neural networks, transformer architectures, self-configuring segmentation systems, post hoc explainability methods and a report-based large language model. Their results could not be pooled because they addressed different tasks, including arterial-phase hyperenhancement, washout, enhancing capsule, ancillary features, LR-3 versus LR-4/5 discrimination, multicategory grading, LR-M subtyping, lesion detection and hepatocellular carcinoma diagnosis. Several studies reported high internal area under the receiver operating characteristic curve or classification accuracy, but the strongest direct category-classification pilot dropped from 60.4% internal accuracy to 41.2% in an external CT cohort and 47.7% in an external MRI cohort.

A stepwise MRI system also performed worse externally on individual major features, particularly washout and capsule, although its category accuracy was relatively stable. More recent multicenter studies reported greater robustness and clinically meaningful reader assistance. However, 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 addressed a different task from image-level recognition.

The review concludes that available evidence supports AI as an aid to standardized feature assessment and reader support, not as a replacement for LI-RADS interpretation. The authors say clinical readiness requires independent multivendor validation, appropriate pathology or rigorous longitudinal reference standards, locked algorithms, calibrated outputs, evaluation of human–AI workflows, and transparent reporting of exclusions, indeterminate cases and domain shift.