Knowledge-guided dual-path AI framework classifies liver MRI series with high accuracy

A knowledge-guided dual-path AI framework achieved macro-average F1-scores of 97.63% (internal) and 96.76% (external) for classifying liver MRI series across 18 categories, significantly outperforming an image-only model (p<0.001).

• The framework integrated CNN imaging features and DICOM metadata with rule-based fusion, achieving macro-average F1-scores of 97.63% (internal, n=7,141 series) and 96.76% (external, n=2,208 series, 22 hospitals).
• On 12 shared categories, the dual-path model significantly outperformed an image-only model (macro F1 96.50% vs 91.54%, McNemar χ²=53.6, p<0.001).
• The model demonstrated robust external validation across multiple centers, showing potential for clinical integration.

Automated summary

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