A CT-Based Deep Learning Radiomics Scoring System for Predicting the Prognosis to Repeat TACE in Patients with...

This study aimed to construct a novel retreatment scoring system to screen patients with hepatocellular carcinoma (HCC) who could benefit further after transarterial chemoembolization (TACE).

310 patients with HCC were retrospectively recruited from three hospitals. The training and validation cohort were randomly selected from Center 1, and two external testing cohorts comprised from Center 2 and Center 3, respectively. Deep learning score and handcrafted radiomics signatures were constructed from the pretreatment arterial-phase and venous-phase CT images. The optimal features were screened using SelectKBest and LASSO regression. The AUC of the optimal combined model, consisting of HBsAg, five radiomics features, and DLscore, was 0.97, 0.89, 0.76, and 0.84 in the four cohorts, respectively. The optimal model was well calibrated. The prediction performance was assessed with respect to receiver operating characteristics, calibration, and decision curve analysis. Kaplan-Meier survival curves based on the scoring system were used to estimate the overall survival (OS).

The optimal combined model consisted of HBsAg, 5 radiomics signatures, and DLscore, which AUC in four cohorts was 0.97, 0.89, 0.76, and 0.84, respectively, with good calibration. Decision curve analysis confirmed that the combined model was clinically useful. After Cox regression analysis of these characteristics, the scoring system (HBsAg-Radscore-DLscore, HRD) was significantly associated with OS in patients with HCC, and was superior to the traditional ART score and ABCR score between high and low-risk patients.

Deep learning and radiomics had good performance in predicting the OS of patients with HCC treated with repeated TACE. The HRD score is a potentially valuable and intelligent prognostic scoring system better than the traditional score.