Interpretable AI supports small liver-lesion assessment on multiphase CT
Published in Artificial Intelligence, Journals, Scanners and Imaging, Tools.
Published October 30, 2026.
Researchers describe dLI-RADS, an interpretable AI method for estimating hepatocellular carcinoma probability from multiphase CT scans. It combines neural-network predictions with quantitative measurements related to established LI-RADS imaging features.
The retrospective multicenter study used training, internal evaluation and external evaluation groups of 143, 62 and 133 patients, respectively. Lesions larger than 40 millimeters were excluded, and liver lesions required manual segmentation.
Internally, lesion-level accuracy was 80% for dLI-RADS and 74% for LI-RADS, meeting the study’s non-inferiority criterion. Externally, sensitivity was 93% versus 78%; dLI-RADS specificity was 74%. Because comparator features were unavailable for non-cancer lesions in that external set, only sensitivity could be directly compared.
These results support further investigation as decision support, not autonomous diagnosis. Non-inferiority does not establish superiority, and retrospective performance does not demonstrate improved patient outcomes. Prospective testing and evaluation of workflow requirements remain necessary.
Source: Sarfati et al., Scientific Reports. Published October 30, 2026. DOI: 10.1038/s41598-026-77611-5.