FIBRONET: AN EXPLAINABLE DEEP LEARNING FRAMEWORK FOR MULTI-STAGE LIVER FIBROSIS CLASSIFICATION FROM ULTRASOUND IMAGING

Abstract

Liver fibrosis is a progressive pathological disorder characterized by the excessive deposition of extracellular matrix proteins as a response to chronic liver injury. If left undetected, it may progress to severe complications such as cirrhosis, liver failure, and hepatocellular carcinoma. Early and accurate diagnosis of fibrosis stages is therefore critical for effective clinical management and treatment planning. Ultrasound imaging is widely utilized for liver disease screening due to its non-invasive nature, low cost, and suitability for continuous clinical monitoring. However, reliable interpretation of ultrasound images remains challenging because of inherent noise, operator dependency, and the subtle structural variations that occur across different stages of fibrosis. To overcome these challenges, this study introduces FibroNet, an explainable deep learning framework designed for automated multi-stage classification of liver fibrosis using ultrasound images. The proposed architecture incorporates a deep convolutional neural network for hierarchical feature extraction, followed by an attention-based feature refinement mechanism that emphasizes clinically significant patterns within ultrasound scans. Furthermore, an explainable artificial intelligence (XAI) component based on Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated to enhance the interpretability of model predictions by visually highlighting diagnostically relevant regions. The model was trained and evaluated on a dataset comprising over 6,000 ultrasound images representing five fibrosis stages (F0–F4). Experimental results demonstrate that FibroNet outperforms several baseline deep learning models, achieving an overall classification accuracy of 94.7% and a ROC–AUC score of 0.97. In addition, the integrated explainability module provides clear visual evidence of the regions contributing to classification decisions, thereby improving transparency and supporting clinical trust in automated predictions.  Overall, the findings suggest that the proposed framework can function as a reliable computer-aided diagnostic system for non-invasive liver fibrosis staging, with strong potential for integration into intelligent healthcare platforms and clinical decision support systems