Study on Ultrasound-Assisted Diagnosis of CHB Complicated with NAFLD Hepatic Fibrosis Based on Deep Learning

This study aimed to develop and evaluate an automated classification model for diagnosing chronic hepatitis B (CHB) complicated with non-alcoholic fatty liver disease (NAFLD) using deep learning techniques applied to two-dimensional liver imaging. We retrospectively analyzed data from 2803 patients diagnosed with CHB and NAFLD via two-dimensional ultrasound and FibroScan at our hospital between June 2019 and December 2022. These patients contributed a total of 20,540 two-dimensional liver images. An additional 150 patients with CHB complicated with NAFLD, who had liver biopsy results, were selected (with a total of 922 liver 2D ultrasound images) for validation of the deep learning model. The diagnostic performance was assessed using sensitivity, specificity, and accuracy metrics for independent diagnoses made by clinicians or with AI assistance. Diagnostic performance was assessed under blinded conditions by physicians of varying expertise, primary, intermediate, and senior clinicians, either independently or assisted by the AI model. We then conducted a statistical analysis of the results, focusing on sensitivity, specificity, and accuracy metrics. The sensitivity, specificity, and accuracy for independent diagnoses made by primary, intermediate, and senior physicians were 24.3%, 29.1%, and 33.8% for sensitivity; 100% for specificity; and 25.3%, 30.0%, and 34.7% for accuracy, respectively. The AI model alone achieved 68.9% sensitivity, 100% specificity, and 69.3% accuracy. When diagnoses were AI-assisted, the sensitivity increased significantly to 73.7% for primary physicians, 73.7% for intermediate physicians, and 75.7% for senior physicians, while specificity remained at 100%, and accuracy was 74.0%, 74.0%, and 76.0%, respectively. The area under the receiver operating characteristic curve (AUC) for AI-independent diagnosis (0.845) was significantly higher than that for primary, intermediate, and senior physicians (0.622, 0.645, and 0.669, respectively; $P<0.001$ ). Similarly, AI-assisted diagnosis AUC values (0.868, 0.868, and 0.878, respectively) surpassed those for independent diagnosis by physicians at each level of expertise ( $P<0.001$ ). In conclusion, the deep learning model based on two-dimensional liver imaging demonstrated feasibility for assisting in the identification of liver fibrosis in patients with CHB and NAFLD. Our results suggest that integrating AI models into the diagnostic process can significantly improve the accuracy of diagnosing liver conditions in this patient population.