A model based on clinico-biochemical characteristics and deep learning features from MR images for assessing necroinflammatory activity in chronic hepatitis B

Accurate liver necroinflammatory activity diagnosis could guide the clinical decision-making in chronic hepatitis B (CHB) patients.This study aimed to build a non-invasive diagnostic model for liver necroinflammatory activity by incorporating deep learning features and clinico-biochemical characteristics in CHB patients. A total of 239 CHB patients who underwent liver biopsy were recruited and randomly divided into a training cohort (n = 179) and an independent validation cohort (n = 60). Bidirectional stepwise selection identified independent clinico-biochemical characteristics. Multivariate logistic regression analysis was used to establish the final combined model by incorporating clinico-biochemical and deep learning features. Predictive performance was evaluated by discrimination and clinical usefulness. Immunoglobulin M, platelets, laminin, type IV collagen, gamma-glutamyl transferase, alanine aminotransferase, aspartate transaminase, alkaline phosphatase, direct bilirubin, and total bilirubin were identified as independent factors. The combined model exhibited better performance than models based on clinico-biochemical characteristics alone, with an AUC of 0.942 (95% confidence interval [CI], 0.912-0.969) for necroinflammatory activity ≥G2 and 0.885 (95% CI, 0.829-0.934) for ≥G3 in the training cohort, and 0.938 (95% CI, 0.867-0.993) and 0.854 (95% CI, 0.764-0.934) in the validation cohort, respectively. The decision curve confirmed its clinical usefulness. The combined model provided an accurate non-invasive prediction of liver necroinflammatory activity, which might contribute to clinical decision-making in CHB patients.