A NOVEL MECHANISM ON FUSION CNN–RADIOMICS APPROACH FOR HEPATITIS LIVER FIBROSIS DETECTION

Abstract

Hepatitis-induced liver fibrosis is a progressive pathological change that, if undiagnosed, can lead to cirrhosis and hepatocellular carcinoma. Early and accurate detection using liver imaging remains a clinical challenge due to subtle textural variations and heterogeneous lesion patterns. This paper proposes a Novel Fusion CNN–Radiomics Framework (FCRF) integrating deep convolutional feature extraction with handcrafted radiomic descriptors for robust classification of liver fibrosis stages. Radiomics features extracted from ultrasound and CT liver regions of interest (ROIs) are combined with a CNN backbone using a feature-level fusion mechanism based on attention-weighted concatenation. The fused feature matrix is optimized using a Sparse Feature Selection Layer and classified using a fully connected deep classifier. Experimental evaluation on a curated liver fibrosis dataset demonstrates that FCRF achieves 94.87% accuracy, 93.12% F1-score, and 0.96 AUC, outperforming CNN-only, radiomics-only, and classical ML baselines. The results confirm that fusion of deep and handcrafted descriptors significantly enhances fibrosis detection performance and supports clinical decision making