Multiparametric ultrasound diagnostics with artificial intelligence programs in diffuse liver diseases in children

Purpose: The aim is to evaluate the diagnostic accuracy of multiparametric ultrasound (US), including shear wave elastography (2D-SWE) and machine learning (ML) models, for the staging of liver fibrosis in children.

Methods/Materials: Materials and methods: a retrospective single-center study included 185 children - mean age 11±3.4 years (140 patients - with verified chronic diffuse liver diseases (HDL), in particular: autoimmune hepatitis, Wilson's disease, chronic viral hepatitis B/C and metabolically associated steatotic liver disease (MASLD) 45 children - control). All patients underwent 2D-SWE, hepatic vein Doppler, and B-mode analysis with two ML models: Model A (Random Forest, binary classification) and Model B (hierarchical ensemble, multiclass classification). Clinical verification and risk stratification by SWE served as the reference method.

Results:

Results: a strong correlation was found between SWE data and Model B forecasts (ρ=0.77). Model A demonstrated a specificity of 84% in differentiating normal from pathology. Model B showed a higher sensitivity (83%), but a tendency to overdiagnosis of the initial stages of fibrosis (reclassification of F0–F1 into F2). Doppler patterns (monophasic spectrum of blood flow in the liver veins ) were reliably associated with higher stages of fibrosis.

Conclusions:

Conclusions: the integration of ML-algorithms into the ultrasound protocol increases the objectivity of diagnosis, however, hierarchical models require calibration to minimize false-positive results in the F1–F2 zone.