Hyperspectral imaging combined with residual-attention-net for spectral-spatial feature fusion in liver disease diagnosis

Distinguishing well-differentiated hepatocellular carcinoma (HCC) from cirrhosis is critical for effective treatment. However, while pathological morphology remains the gold standard, it has limitations in differentiating these two conditions. This study aims to propose a novel hyperspectral image (400–1000 nm) processing method based on 3D-Residual-attention networks (3D Ra-Net) to improve the accuracy of differentiation between the two.The study employs a 3D Ra-Net model that integrates spectral features with spatial information to enhance classification accuracy. We incorporated band selection techniques, including the Norris derivative and the Successive Projections Algorithm (SPA), and optimized the data processing workflow. Experimental performance was evaluated using cross-validation, with the primary metrics of accuracy, sensitivity, and specificity for statistical analysis. The experimental results demonstrate that the 3D Ra-Net model achieved a classification accuracy of 92.11 % in distinguishing well-differentiated HCC from cirrhosis. Additionally, the model achieved an accuracy of 84.67 % in distinguishing well-differentiated HCC, poorly differentiated HCC, cirrhosis, and normal liver tissue. Sensitivity and specificity values also indicated strong diagnostic performance. The key innovation of this study lies in the development of the 3D Ra-Net model and the efficient extraction of joint spatial-spectral features. This method provides a novel, effective approach for the accurate diagnosis of HCC, offering reliable potential for clinical application in liver disease diagnosis.