Abstract
BACKGROUND AND PURPOSE: Predicting hepatocellular carcinoma (HCC) response to Stereotactic Body Radiation Therapy (SBRT) can be challenging. Here, we assessed the value of a radiomics-based machine learning (ML) approach for predicting HCC response to SBRT, using pre-treatment and early post-treatment magnetic resonance imaging (MRI). MATERIALS AND METHODS: This retrospective single-center study included 87 patients (M 67, mean age 65.3 ± 9.1y) with HCC treated with SBRT who underwent gadoxetate MRI both pre- and early post-treatment (around 9.5 weeks). Tumor radiomics features were extracted on pre- and post-SBRT MRIs on pre- and post-contrast T1-weighted imaging (T1WI) [pre-contrast, arterial phase (AP), portal venous phase (PVP), transitional phase and hepatobiliary phase]. Long term response was assessed using modified RECIST criteria. Different ML models were developed based on 1st and 2nd order radiomics features to predict long-term objective response (partial and complete response) versus no response (stable and progressive disease). The cohort was randomly divided into training/validation (70 %) and testing 30 %. RESULTS: A total of 87 tumors were assessed (mean size 2.7 ± 1.6 cm). Objective long-term response was observed in 43 (49.4 %) patients. The best predictive outcomes were achieved using models combining pre- and early post-treatment radiomics, with top performing model combining pre-treatment T1WI-pre-contrast, pre-treatment T1WI-AP and post-treatment T1WI-PVP, achieving an AUC of 0.85 [95 % CI: 0.67---1], sensitivity of 0.7 and specificity of 1. CONCLUSIONS: Our initial findings show promising results for ML radiomics in predicting long-term response of HCC to SBRT, which may have implications for management decisions.