Objective:

To develop and validate a multiphasic CT-based fusion model integrating deep learning predictions with quantitative enhancement features to differentiate pediatric hepatoblastoma (HB) from focal nodular hyperplasia (FNH).

Approach:

• Study Design: Retrospective analysis of 612 children who received pretreatment multiphasic CT, with 523 having HB and 89 having FNH.
• Model Development: Three phase-specific deep learning models were trained on precontrast, arterial-phase, or portal venous-phase images, and a logistic regression model was developed using quantitative enhancement features.
• Evaluation: The model's performance was evaluated using area under the receiver operating characteristic curve (AUC), diagnostic accuracy measures, and comparison against radiologist interpretation.

Key Findings:

• The DL+QEF model achieved an AUC of 0.980 and 94.2% accuracy in the temporal test cohort.
• Sensitivity for HB was 94.9% and specificity for FNH was 92.9%.
• Junior-radiologist accuracy improved from 88.4% to 95.0% with model assistance.

Interpretation:

The findings suggest that the DL+QEF model significantly enhances the differentiation between HB and FNH in pediatric patients, potentially leading to improved diagnostic accuracy and treatment planning.

Limitations:

• The study is retrospective and may not fully capture all clinical scenarios.
• The model's performance in real-world settings outside the study cohort is not established.

Conclusion:

The integration of deep learning with quantitative enhancement features in a multiphasic CT framework shows promise in differentiating pediatric HB from FNH, warranting further validation in diverse clinical settings to confirm its utility in routine practice.