New deep learning model classifies liver fibrosis stages from ultrasound with 99% accuracy

Researchers have developed CLAUNet-LFB0-CBAM, a deep learning framework that classifies liver fibrosis stages from ultrasound images with 99.05% accuracy, outperforming standard models. The hybrid architecture combines an augmented EfficientNet-B0 network with CBAM attention mechanisms, using weighted loss functions and sampling to address class imbalance. It achieved a 0.980 F1-score and 0.9987 AUC, with notable improvements in intermediate fibrosis stages critical for clinical assessment. Grad-CAM analysis confirmed the network learns from relevant anatomical liver regions. The model is lightweight at approximately 4.8 million parameters. The research was conducted at VIT-AP University and published under a Creative Commons license.