Examining the Effects of Slice Thickness on the Reproducibility of CT Radiomics for Patients with Colorectal Liver Metastases

Abstract

We present an analysis of 81 patients with colorectal liver metastases from two major cancer centers prospectively enrolled in an imaging trial to assess reproducibility of radiomic features in contrast-enhanced CT. All scans were reconstructed with different slice thicknesses and levels of iterative reconstruction. Radiomic features were extracted from the liver parenchyma and largest metastasis from each reconstruction, using different levels of resampling and methods of feature aggregation. The prognostic value of reproducible features was tested using Cox proportional hazards to model overall survival in an independent, public data set of 197 hepatic resection patients with colorectal liver metastases. Our results show that larger differences in slice thickness reduced the concordance of features ($p<10^{-6}$). Extracting features with 2.5D aggregation and no axial resampling produced the most robust features, and the best test-set performance in the survival model on the independent data set (C-index = 0.65). Across all feature extraction methods, restricting the survival models to use reproducible features had no statistically significant effect on the test set performance ($p=0.98$). In conclusion, our results show that feature extraction settings can positively impact the robustness of radiomics features to variations in slice thickness, without negatively effecting prognostic performance.