Multicenter validation of an interpretable ML model on Gd-EOB-DTPA MRI for differentiating small (≤2 cm) HCC from benign lesions

Gd-EOB-DTPA-MRI fusion ML model for small HCC (≤2 cm) yielded internal AUC 0.913, external AUC 0.824, comparable to senior radiologists; improvement over radiomics-only not significant.

• The fusion model integrated radiomics, qualitative MRI features, and clinical data; SHAP identified HBV status, AFP, and hepatobiliary phase radiomics as top predictors.
• In external validation, AUC remained stable after ComBat harmonization (AUC 0.806), suggesting robustness to scanner variability.
• The numerical advantage over junior radiologists did not reach statistical significance due to limited external sample size.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.