Ultrasound - based deep learning radiomics nomogram for noninvasive prediction of p53 mutation status in hepatocellular carcinoma: a variational autoencoder based development and validation study.

Objectives: Accurate preoperative assessment of p53 mutation status in hepatocellular carcinoma (HCC) is critical for prognostic stratification and personalized treatment planning. Conventional radiomics approaches often suffer from feature redundancy and limited generalization. This study aimed to develop and validate a noninvasive ultrasound-based radiomics nomogram integrating variational autoencoder (VAE)-derived deep features to predict p53 expression status, addressing these limitations. Methods: A retrospective cohort of 172 patients with pathologically confirmed HCC (training cohort: n=120, validation cohort: n=52) who underwent preoperative two-dimensional ultrasound images and had available p53 immunohistochemistry (IHC) results was analyzed. ultrasound images were segmented, and radiomic features were extracted from them. A VAE network was employed to reduce feature dimensionality and extract high-level malignant risk scores. These scores were integrated with clinical variables (Alpha-Fetoprotein [AFP] levels, Microvascular Invasion [MVI] status, and Edmondson-Steiner (E-S) grade) to construct a predictive nomogram. Model performance was evaluated using receiver operating characteristic (ROC) analysis (area under the curve [AUC]), calibration curves, and decision curve analysis (DCA). Results: The VAE-integrated nomogram achieved robust predictive performance, with an AUC of 0.925 (95% CI: 0.881-0.969) in the training cohort and 0.820 (95% CI: 0.699-0.942) in the validation cohort. Calibration curves demonstrated close alignment between predicted and observed probabilities, and decision curve analysis confirmed clinical utility across a broad threshold probability range. Key clinical benefits included noninvasive assessment of p53 mutation and enhanced interpretability through combined deep learning and clinical parameters. Conclusion: This VAE-based radiomics framework effectively combines deep feature representation with clinical variables, providing a reliable tool for noninvasive preoperative evaluation of HCC p53 mutation. The model shows promise for optimizing surgical decision-making and personalized prognostic strategies in HCC management.