A pretreatment multiphasic CT-based decision-support model for differentiating pediatric hepatoblastoma from focal nodular hyperplasia.

Background: Hepatoblastoma (HB) and focal nodular hyperplasia (FNH) require markedly different management strategies in children, but differentiation can be challenging when multiphasic computed tomography (CT) findings are atypical or overlapping. This study aimed to develop and temporally validate a pretreatment multiphasic CT-based fusion model combining deep learning (DL) predictions with quantitative enhancement features for pediatric HB-FNH differentiation and to evaluate the potential value of model assistance for junior-radiologist interpretation. Methods: This retrospective study included 612 children who underwent pretreatment multiphasic CT between January 2011 and September 2025, including 523 with HB and 89 with FNH. A chronological split on July 1, 2023 assigned 491 patients to model development and 121 patients to temporal testing. Three phase-specific DL models based on 2.5D ResNet-18 were trained using precontrast, arterial-phase, and portal venous-phase images, and their predicted probabilities were averaged to generate DL-3Phase. Quantitative enhancement features from standardized manual region-of-interest attenuation measurements were used to train a regularized logistic regression model, termed the quantitative enhancement feature (QEF) model. DL+QEF was prespecified as the unweighted average of the DL-3Phase and QEF model probabilities. Performance was assessed using the area under the receiver operating characteristic curve (AUC), diagnostic accuracy metrics, calibration, decision curve analysis, and comparison with radiologist interpretation. Results: In the temporal test cohort, DL+QEF achieved an AUC of 0.980, an accuracy of 94.2%, an HB sensitivity of 94.9%, and an FNH specificity of 92.9%. DL+QEF showed better discrimination than DL-3Phase and the best-performing single-phase DL model, and combined the high HB sensitivity of DL-3Phase with the high FNH specificity of the QEF model. In the pathology-confirmed test subset, the AUC was 0.972 and accuracy was 92.9%. Junior-radiologist accuracy was 88.4% during unaided reading and 95.0% during model-assisted reading (exact McNemar Conclusion: The pretreatment multiphasic CT-based DL+QEF model showed favorable diagnostic performance for pediatric HB-FNH differentiation and may serve as an adjunct to CT interpretation, with potential utility in reader-defined atypical or diagnostically uncertain cases.