Abstract:
Liver fibrosis represents a gradual replacement of normal hepatic parenchyma with scar tissue, and accurate staging (F0–F4) is essential for guiding surveillance, therapeutic decisions, and specialist referral. Although biopsy remains the historical gold standard, it is invasive, costly, and prone to sampling variability. Ultrasound (US), by contrast, is non- invasive, broadly accessible, and suitable for repeated monitoring in chronic liver disease; however, intrinsic challenges such as depth-dependent attenuation, speckle noise, and inter-scanner heterogeneity complicate automated staging. In this study, developed an explainable semi-supervised learning framework for five-stage fibrosis classification using heterogeneous US data from tertiary-care hospitals, capturing real-world class imbalance with predominance of F0 and F4 cases. The framework integrates Mean Teacher consistency learning with prototypical loss to enhance stage-aware embeddings, while a class-balanced focal loss addresses long-tailed distributions. Under limited labeling budgets (5–15% per class), the approach achieves robust performance (test accuracy 93.46%, macro-F1 91.86%, κ = 0.945, ROC-AUC 0.999), with Grad-CAM highlighting clinically meaningful intraparenchymal features. These results demonstrate the potential of a prototype-regularized semi-supervised pipeline to deliver accurate, interpretable fibrosis staging from routine US, while markedly reducing annotation demands. Future directions include cross-center validation, calibration, and prospective clinical evaluation.
