Room: Davidson Ballroom B
Purpose: About half of all hepatocellular carcinoma (HCC) patients treated with SBRT fail regionally. This work looks to identify a CT-based radiomics signature associated with liver regional progression, and to reveal underlying imaging biomarkers which promise to support clinical decision making.
Methods: 106 HCC patients (59 regional progression) were retrospectively analyzed. For each, 3441 radiomic features were extracted for both the gross tumor volume (GTV) and Liver-GTV regions from pretreatment contrast-enhanced CT images with varying sampling and quantization parameters: 5 image resolutions (1-5 mm), 4 gray scale levels (8, 16, 32, 64) and 2 quantization methods (Lloyd and equal probability). The model building process consisted of feature set reduction, feature selection and prediction performance evaluation. A graph-based feature ranking method, which considers feature relevance and variance, was used to reduce the candidate set to 25 features. A radial basis function (RBF) based support vector machine was applied for model building. Five-fold cross-validation (CV) was used to obtain optimal RBF width, regularization parameter (C) and model order during training. Prediction performance was evaluated on independent test data (46 HCC patients, 15 events) and assessed by the area under the receiver-operating characteristic curve (AUC).
Results: The resulting model contains 7 predictive radiomic features from both GTV and Liver-GTV regions, 3 features from the gray level size zone matrix (GLSZM), 3 from gray level run length matrix (GLRLM), and one from neighborhood gray tone difference matrix (NGTDM). The model reached an AUC of 0.73 (CI: 0.58, 0.81)) on CV during training and an AUC of 0.72 on the test set.
Conclusion: A radiomics signature from contrast enhanced CT images shows promise for assessing liver regional progression risk, a potentially ground-breaking step to improve the prognosis and treatment management of HCC patients.