Multiparametric MRI based deep learning model for prediction of early recurrence of hepatocellular carcinoma after SR following TACE.

backgroundSurgical resection (SR) following transarterial chemoembolization (TACE) is a promising treatment for unresectable hepatocellular carcinoma (uHCC). However, biomarkers for the prediction of postoperative recurrence are needed. PURPOSE: To develop and validate a model combining deep learning (DL) and clinical data for early recurrence (ER) in uHCC patients after TACE.

methodsA total of 511 patients who received SR following TACE were assigned to derivation (n = 413) and validation (n = 98) cohorts. Deep learning features were taken from the largest tumor area in liver MRI. A nomogram using DL signatures and clinical data was made to forecast early recurrence risk in uHCC patients. Model performance was evaluated using area under the curve (AUC).

resultsA total of 2278 subsequences and 31,346 slices multiparametric MRI including contrast-enhanced T1WI, T2WI and DWI were input in the DL model simultaneously. Multivariable analysis identified three independent predictors for the development of the nomogram: tumor number (hazard ratio [HR]:3.42, 95% confidence interval [CI]: 2.75-4.31, P = 0.003), microvascular invasion (HR: 9.21, 6.24-32.14; P < 0.001), and DL scores (HR: 17.46, 95% CI: 12.94-23.57, P < 0.001). The AUC of the nomogram was 0.872 and 0.862 in two cohorts, significantly outperforming single-subsequence-based DL mode and clinical model (all, P < 0.001). The nomogram provided two risk strata for cumulative overall survival in two cohorts, showing significant statistical results (P < 0.001).

conclusionsThe DL-based nomogram is essential to identify patients with uHCC suitable for treatment with SR following TACE and may potentially benefit personalized decision-making.