Deep learning combines MRI texture and clinical factors to predict early efficacy in TACE for hepatocellular carcinoma

Abstract Objectives Hepatocellular carcinoma (HCC) remains a leading cause of cancer-related mortality worldwide, and transarterial chemoembolization (TACE) represents the standard treatment for patients with intermediate-stage disease. However, early prediction of post-TACE response remains challenging due to tumor heterogeneity and limited interpretability of existing models. This study aimed to develop and validate a deep learning (DL) framework that integrates multi-sequence MRI texture features with preoperative clinical factors to predict early postoperative efficacy following TACE. Methods In this multicenter retrospective study, 178 HCC patients who underwent TACE were enrolled from two institutions. Dynamic contrast-enhanced MRI data and clinical variables were analyzed to construct a Transformer-based DL model incorporating a multi-head attention mechanism for sequence-level feature integration. The model’s prediction probability and significant clinical factors were further combined through logistic regression to generate a nomogram for individualized response estimation. Results The DL model achieved robust predictive performance, with an area under the curve (AUC) of 0.90 in the internal testing set and 0.79 in the external validation set. Incorporating clinical variables improved predictive accuracy, yielding AUCs of 0.98 and 0.86 for the in-house and external datasets, respectively. Calibration and bootstrap analyses confirmed model stability, while decision curve analysis demonstrated superior clinical utility compared with conventional approaches. Attention-weight analysis revealed that T2-weighted and diffusion-weighted sequences contributed most to prediction, consistent with their biological relevance in tumor viability assessment. Conclusions The integrated DL nomogram provides an interpretable, noninvasive tool for predicting early TACE response in HCC, potentially aiding in personalized therapeutic decision-making and optimizing treatment strategies.