Efficacy of Radiomics in Predicting Oncologic Outcome of Liver-Directed Combined Radiotherapy in Locally Advanced Hepatocellular Carcinoma

Abstract

Simple Summary In hepatocellular carcinoma (HCC), the clinical predictive factors for tumor markers are well-known. Although these factors are recognized as essential, recent attempts have been made to predict treatment outcomes using radiomics based on imaging markers. We investigated whether radiomic features extracted from three-phase dynamic contrast-enhanced computed tomography (CECT) can be used to predict clinical outcomes, including objective treatment response (OR) and in-field failure-free survival rate (IFFR), in 409 patients with HCC who received liver-directed combined radiotherapy (LD-CRT). In predicting the OR and IFFR, clinical models and radiomics models based on tumoral and peritumoral areas showed an acceptable performance, while combined clinico-radiomics models (CCR) performed better. Therefore, CCR models have potential use in clinical prediction. Moreover, the constructed nomograms based on these models may provide valuable information on the OR and IFFR in patients with HCC undergoing LD-CRT.Abstract Purpose: We investigated whether radiomic features extracted from three-phase dynamic contrast-enhanced computed tomography (CECT) can be used to predict clinical outcomes, including objective treatment response (OR) and in-field failure-free survival rate (IFFR), in patients with hepatocellular carcinoma (HCC) who received liver-directed combined radiotherapy (LD-CRT). Methods: We included 409 patients, and they were randomly divided into training (n = 307) and validation (n = 102) cohorts. For radiomics models, we extracted 116 radiomic features from the region of interest on the CECT images. Significant clinical prognostic factors are identified to predict the OR and IFFR in the clinical models. We developed clinical models, radiomics models, and a combination of both features (CCR model). Results: Among the radiomic models evaluated for OR, the OR-PVP-Peri-1cm model showed favorable predictive performance with an area under the curve (AUC) of 0.647. The clinical model showed an AUC of 0.729, whereas the CCR model showed better performance (AUC 0.759). For the IFFR, the IFFR-PVP-Peri-1cm model showed an AUC of 0.673, clinical model showed 0.687, and the CCR model showed 0.736. We also developed and validated a prognostic nomogram based on CCR models. Conclusion: In predicting the OR and IFFR in patients with HCC undergoing LD-CRT, CCR models performed better than clinical and radiomics models. Moreover, the constructed nomograms based on these models may provide valuable information on the prognosis of these patients.