Liver fibrosis leads to progressive liver scarring, and accurate staging from F0 to F4 is essential for patient monitoring, treatment planning, and clinical management. Although ultrasound is a noninvasive, widely accessible imaging modality, reliable fibrosis staging remains difficult, particularly at intermediate stages. Existing machine learning approaches are further limited by the scarcity of labeled data and by largely post hoc interpretability methods. This study proposes an explainable semi-supervised learning framework that combines a Mean Teacher strategy with class-balanced loss and prototype regularization to enable robust learning from limited annotations. A broad supervised benchmark is constructed using representative convolutional and transformer-based networks to establish strong reference performance. The proposed framework consistently outperforms these baselines, achieving high classification accuracy and agreement across fibrosis stages on a real ultrasound dataset with only a small fraction of labeled samples. In addition, the model produces well-calibrated predictions and demonstrates clinical utility for identifying advanced fibrosis. Most classification errors occur between adjacent stages, reflecting the continuous nature of disease progression, and visual explanations highlight periportal and parenchymal regions known to be relevant for staging. Overall, the framework provides accurate, reliable, and interpretable fibrosis assessment while substantially reducing annotation requirements. • Deliver a semi-supervised method that reduces the need for extensive labeled ultrasound data. • Improve recognition of subtle fibrosis stages through stable prototypical learning. • Strengthen model performance with balanced loss functions and tailored data augmentation. • Provide interpretable outputs using heatmaps and prototype-based visual cues. • Support clinical decision making with calibrated predictions for fibrosis staging.
Shakil et al. (Wed,) studied this question.