Radiomic analysis on pretreatment MRI to predict response to atezolizumab plus bevacizumab in advanced hepatocellular carcinoma: A multicenter study.

Abstract

4104 Background: The combination of atezolizumab and bevacizumab is the standard treatment for advanced hepatocellular carcinoma approved in China and many other countries. However, the objective response rate of this combination treatment was around 30%. Consequently, identifying individuals with the potential to respond favorably prior to initiating therapy remains a pressing challenge. Methods: This multi-center retrospective study included advanced hepatocellular carcinoma patients who received atezolizumab plus bevacizumab as first-line therapy between December 2020 and February 2024. The training cohort consisted of eligible patients who have complete baseline, treatment and tumor evaluation records and MRI imaging data from Zhongshan Hospital, while eligible patients from other centers constituted the external validation cohort. A deep learning model based on nnU-Net was developed to automatically segment intrahepatic lesions. All segmentations were reviewed and revised by two radiologists. The radiomic features were extracted using PyRadiomics, then a radiomic feature-based model for predicting response to atezolizumab plus bevacizumab therapy was constructed using the Extreme Gradient Boosting Decision Tree (XGBoost) algorithm. Additionally, three radiologists evaluated 53 visually-assessed MRI features on MRI scans. Finally, the predictive performance of radiomic feature model, as well as the relationship between radiomic and MRI features, was assessed. Results: A total of 240 eligible patients were recruited from 14 centers in China, of which 161 and 79 were classified as training and validation cohorts, respectively. During a median follow-up period of 13.7 months (IQR: 8.3–20.6) in the training cohort and 10.5 months (IQR: 7.4–17.2) in the validation cohort, 19.0% (30/161) and 23.0% (18/79) of patients, respectively, achieved an objective response by RECIST v1.1 (p = 0.559). The radiomic feature model demonstrated a promising predictive performance, achieving an AUC of 0.913 (95% CI: 0.874–0.953) in the training cohort and 0.825 (95% CI: 0.700–0.949) in the validation cohort. Fat surpassing liver mass was the only MRI feature associated with an objective response (p = 0.020). When the MRI feature was combined with radiomic features, the predictive model further improved, yielding an AUC of 0.951 (95% CI: 0.924–0.979) in the training cohort and 0.835 (95% CI: 0.725–0.945) in the validation cohort. A significant correlation was observed between radiomic features and MRI features of intrahepatic lesions with a univariate analysis p < 0.2. Conclusions: Radiomic features derived from pretreatment MRI scans can effectively predict personalized objective responses to combination therapy with atezolizumab and bevacizumab in patients with unresectable or advanced HCC.