Deep Learning Model Predicts Hepatocellular Carcinoma Progression Following Transarterial Chemoembolization

• A multicenter study of 508 patients with unresectable hepatocellular carcinoma (HCC) evaluated a deep learning–based model for predicting time to progression following transarterial chemoembolization (TACE).
• The model integrated deep learning features from pretreatment arterial-phase CT images with radiomics and clinical data, demonstrating consistent predictive performance across internal and external validation cohorts.
• Further prospective validation is needed to determine whether the model can improve patient selection and individualized treatment planning.

A new prediction model combining deep learning, radiomics, and clinical data demonstrated promising performance in predicting disease progression following transarterial chemoembolization (TACE) in patients with hepatocellular carcinoma (HCC). The multicenter study by Wang and colleagues was published September 30, 2026, in Cancer Imaging.¹

To predict the time to progression (TTP) of HCC after TACE, investigators developed a prediction model, TTP-Net, by integrating clinical variables, radiomics, and 2.5-dimensional deep learning features extracted from pretreatment arterial-phase CT images. Four other models were developed for comparison, based on the following: (1) the selected clinical variables, (2) the selected radiomics features, (3) deep features extracted from the largest tumor area, and (4) deep features from the 2.5D arterial-phase computed tomography (CT) images.

The retrospective study included 508 patients with treatment-naive, unresectable HCC who underwent conventional TACE as their initial treatment. Two hundred twenty-six patients were included in the public dataset, which was randomly split into training and internal-validation cohorts, and 282 patients were included in 2 independent external-validation cohorts. TTP was the primary endpoint.

The model achieved concordance indices of 0.719 and 0.723 in the training and internal-validation cohorts, respectively, and 0.713 and 0.707 in the external-validation cohorts. Corresponding 12-month areas under the receiver operating characteristic curve were 0.872, 0.862, 0.816, and 0.832. TTP-Net also significantly distinguished between patients at higher and lower risk of disease progression across all cohorts, and decision-curve analysis suggested that TTP-Net may offer greater clinical utility than the comparator models for predicting 12-month disease progression across a wide range of risk thresholds in most cohorts.

The findings suggest that integrating deep learning–derived imaging features with radiomics and clinical data can help identify patients at increased risk of early progression following TACE, addressing an important gap in pretreatment risk stratification. Such information could support a move toward individualized surveillance and treatment planning for patients with unresectable HCC.

However, the retrospective findings require prospective validation. Although TTP-Net demonstrated consistent predictive performance across multiple cohorts, the study did not establish whether using the model to guide clinical decisions improves patient outcomes. Additionally, the study only compared TTP-Net with the 4 models developed by the investigators, not with conventional radiological assessment.

This multicenter study suggests that combining deep learning, radiomics, and clinical data may improve prediction of HCC progression following TACE compared with individual prediction models and help identify patients at increased risk of early progression. Larger prospective studies are needed to validate these findings and establish the model's clinical utility.

1. Wang L, Xia C, Peng Z, et al. Pretreatment arterial-phase CT-based prediction of time to progression after transarterial chemoembolization in hepatocellular carcinoma: a multicenter study. Cancer Imaging. 2026. doi:10.1186/s40644-026-01136-3