Baseline and Longitudinal CT Radiomics of Liver Parenchyma during Image-Guided Radiotherapy Are Associated with Liver Functional Decline and Prognosis
Abstract
Purpose
The sensitivity of normal liver tissue raises concerns about potential liver function post-cancer treatment. We investigate whether liver radiomics features derived from standard CT-guided radiation therapy (RT) can improve acute functional liver decline prediction.
Methods
146 liver cancer patients (66±11 years, n=105 male) treated with hypofractionated RT were analyzed retrospectively. Tumor types included hepatocellular carcinoma (n=40), cholangiocarcinoma (n=72), metastasis (n=27). Liver function decline was defined as any of the following within four months post-treatment: increase in Child-Pugh score by >2 points; alkaline phosphatase levels >2x the upper limit of normal (ULN) or pre-treatment value; liver transaminase levels >5x ULN or pre-treatment value. Each patient underwent a pre-treatment contrast-enhanced planning CT, and daily non-contrast CT-on-rails scan for image-guided radiation. Patients underwent a complete blood panel before and after treatment. The CTs were registered to the planning CT using a biomechanical deformable image registration model, enabling cross-fractional dose accumulation. Radiomics feature changes—within 5Gy accumulated dose bins—between the first and middle treatment fractions were input features in predictive models trained with data available at the pre-treatment and mid-treatment time points. Baseline radiomics were extracted from the planning CT. Gradient Boosting (GB) and Random Forest (RF) models were trained with clinical features—including blood values, tumor type, age—and with/without imaging features. Area under the receiving operator curve (AUC) assessed model performance in a 70/10/20 train/validation/test split.
Results
Mid-treatment imaging features improved model performance compared to baseline imaging. Mid-treatmentimaging had an AUC=0.86 (GB). Pre-treatmentimaging (RF) had an AUC=0.80, versus AUC=0.76 (GB, clinical features only). The salient features were blood biomarkers and texture heterogeneity change within dose bins.
Conclusion
Radiomics feature changes between the first-seventh fractions from CT-on-rails imaging improve liver functional decline prediction, which may help identify patients at-risk for poor outcomes. Ongoing work will investigate the functional declines’ clinical context.