Machine learning-based analysis of CT radiomics model for prediction of colorectal metachronous liver metastases

Purpose Early identification of patients at risk of developing colorectal liver metastases can help personalizing treatment and improve oncological outcome. The aim of this study was to investigate in patients with colorectal cancer (CRC) whether a machine learning-based radiomics model can predict the occurrence of metachronous metastases. Methods In this multicentre study, the primary staging portal venous phase CT of 91 CRC patients were retrospectively analysed. Two groups were assessed: patients without liver metastases at primary staging, or during follow-up of >= 24 months (n = 67) and patients without liver metastases at primary staging but developed metachronous liver metastases < 24 months after primary staging (n = 24). After liver parenchyma segmentation, 1767 radiomics features were extracted for each patient. Three predictive models were constructed based on (1) radiomics features, (2) clinical features and (3) a combination of clinical and radiomics features. Stability of features across hospitals was assessed by the Kruskal-Wallis test and inter-correlated features were removed if their correlation coefficient was higher than 0.9. Bayesian-optimized random forest with wrapper feature selection was used for prediction models. Results The three predictive models that included radiomics features, clinical features and a combination of radiomics with clinical features resulted in an AUC in the validation cohort of 86% (95%CI 85-87%), 71% (95%CI 69-72%) and 86% (95% CI 85-87%), respectively. Conclusion A machine learning-based radiomics analysis of routine clinical CT imaging at primary staging can provide valuable biomarkers to identify patients at high risk for developing colorectal liver metastases.

Automated summary

A machine learning-based radiomics model was able to predict the occurrence of colorectal liver metastases in patients with colorectal cancer and provide valuable biomarkers for identifying high-risk patients. The predictive models including radiomics features, clinical features, and a combination of both showed promising results with AUC values of 86%, 71%, and 86% in the validation cohort, respectively.