All patients (n = 70) who were treated at the Second Affiliated Hospital, Zhejiang University School of Medicine from December 2015 to June 2017 were included in the study. Treatment and data analysis were conducted according to the Declaration of Helsinki. Ethical approval for retrospective data analysis was obtained from the local ethics committee. The diagnosis of liver cancer was based on the guidelines of the American Association for the study of liver diseases . Portal vein invasion can be determined by the presence of filling defects in a low attenuation cavity near the primary tumor, which can be distinguished on enhanced computed tomography (CT).
In this study, patients received SBRT according to the following criteria: (1) tumor thrombus involving the main portal vein and/or the first portal vein, which was deemed unsuitable for surgery or transarterial chemoembolization; (2) Eastern Cooperative Oncology Group (ECOG) performance status (PS) score of 0–1; (3) absence of refractory ascites; (4) Child–Pugh class A, B, and C; (5) no previous history of radiotherapy for the liver; and (6) availability of more than 700 cc of unaffected liver.
The gross tumor volume (GTV) represents the extent of tumor thrombosis visualized on contrast-enhanced CT and magnetic resonance imaging (MRI). If the extent of primary liver disease is small (less than 5 cm) and adjacent to the PVTT, both are considered to be a part of the GTV. A total dose of 25–50 Gy was prescribed in five fractions over 5–7 days based on the GTV. SBRT plans were generated using the Varian radiation treatment planning system (Eclipse software, Varian Medical Systems, Palo Alto, CA). Treatment was delivered with a Varian Trilogy linear accelerator (Varian Medical Systems, Palo Alto, CA) using a 6-MV photon beam.
The cutoff date for the last follow-up was February 28, 2018 for censored data analysis. The OS was calculated from the start of SBRT to the date of death or the last follow-up visit.
The entire image used for radiomic analysis was obtained from the CT scan acquired prior to SBRT. Enhanced CT imaging was performed using a LightSpeed RT 16 scanner (GE, USA), with 0.25-cm thick slices, which were analyzed to extract the radiomic features from the GTV that contributed to the SBRT plans. Feature extraction was based on the three-dimensional (3D) slicer platform, and performed using the pyradiomics package, which is available at: http://PyRadiomics.readthedocs.io/en/latest/ (accessed on June 30, 2019) .
All statistical analyses were performed using R software, version 3.6.3 (R Foundation for Statistical Computing, Vienna, Austria) and X-tile software, version 3.6.1 (Yale University School of Medicine, New Haven, Conn). Least absolute shrinkage and selection operator (LASSO) regression modeling was used for data dimension reduction, feature selection, and radiomic feature building. Multivariate Cox-regression hazard models were built for the survival outcome, radiomic features, and clinical characteristics presented with the nomogram. The area under the curve (AUC) of the receiver operating characteristic (ROC) curve was used to evaluate the model. The radiomic scores (Rad-scores) were calculated for each patient using a linear combination of selected radiomic features, weighted by their respective coefficients. The cutoff value of the Rad-score was calculated using X-tile software to categorize patients into the high-risk group or low-risk group.