Prognostication of colorectal cancer liver metastasis by CE-based radiomics and machine learning

•Radiomics signature based on contrast-enhanced CT for prediction of disease-freesurvival in colorectal cancer patients with liver metastasis can be constructed withmachine learning algorithms.•The prognostic role of both radiomics signatures can be demonstrated with Coxregression and Kaplan-Meier analyses by external validation.•Radiomics signatures by two different machine learning algorithms show comparable stratification performance. The liver is the most common organ for the formation of colorectal cancer metastasis. Non-invasive prognostication of colorectal cancer liver metastasis (CRLM) may better inform clinicians for decision-making. Contrast-enhanced computed tomography images of 180 CRLM cases were included in the final analyses. Radiomics features, including shape, first-order, wavelet, and texture, were extracted with Pyradiomics, followed by feature engineering by penalized Cox regression. Radiomics signatures were constructed for disease-free survival (DFS) by both elastic net (EN) and random survival forest (RSF) algorithms. The prognostic potential of the radiomics signatures was demonstrated by Kaplan-Meier curves and multivariate Cox regression. 11 radiomics features were selected for prognostic modelling for the EN algorithm, with 835 features for the RSF algorithm. Survival heatmap indicates a negative correlation between EN or RSF risk scores and DFS. Radiomics signature by EN algorithm successfully separates DFS of high-risk and low-risk cases in the training dataset (log-rank test: p < 0.01, hazard ratio: 1.45 (1.07–1.96), p < 0.01) and test dataset (hazard ratio: 1.89 (1.17–3.04), p < 0.05). RSF algorithm shows a better prognostic implication potential for DFS in the training dataset (log-rank test: p < 0.001, hazard ratio: 2.54 (1.80–3.61), p < 0.0001) and test dataset (log-rank test: p < 0.05, hazard ratio: 1.84 (1.15–2.96), p < 0.05). Radiomics features have the potential for the prediction of DFS in CRLM cases.