CT-based deep learning prediction of complete response in intermediate-stage hepatocellular carcinoma treated with drug-eluting beads transarterial chemoembolization

Objective to develop a CT-based deep learning model to predict complete response to drug-eluting beads-transarterial chemoembolization in patients with naive BCLC B hepatocellular carcinoma. Methods This dual-centre retrospective study included 93 patients with BCLC B hepatocellular carcinoma treated with drug-eluting beads-transarterial chemoembolization (median size of 40 mm and 37 mm at Institutions 1 and 2). Complete response was defined as per mRECIST on liver contrast-enhanced CT within two months of treatment. A twin-network deep learning model with spatio-temporal ViViT architecture was developed to predict complete response using baseline dedicated liver CT. The model was extensively trained/tested based on an 8-fold cross-validation approach with an ensemble technique, a model vote system where the outcome is the average of multiple model predictions. Results The complete response rate was 36% (18/50) and 22% (11/49) at Institutions 1 and 2. The model showed high specificity and AUC, as well as moderate sensitivity and balanced accuracy when using either the late arterial phase (0.91 ± 0.12​, 0.86 ± 0.16, 0.43 ± 0.23, ​and 0.67 ± 0.13, respectively) or the portal venous phase (0.90 ± 0.15, 0.85 ± 0.17​,​ 0.57 ± 0.30​, and 0.74 ± 0.16​, respectively). Conclusion The developed CT-based deep learning model predicted complete response to drug-eluting beads-transarterial chemoembolization in patients with naive BCLC B hepatocellular carcinoma with high specificity. It should be refined to improve sensitivity. Advances in knowledge The study underscores the potential of artificial intelligence in precision medicine in patients with hepatocellular carcinoma. While the model shows promise, further research with larger datasets and prospective studies is needed to enhance its predictive power and clinical applicability.