Machine learning for CT texture analysis in chemosensitivity of colorectal liver metastases: Initial results.

Abstract

e15148 Background: Despite advances in cancer treatment over the last decades, more efficacious biomarkers are needed in patients with metastatic colorectal cancer. Several studies have reported that CT texture analysis is a useful prognostic biomarker for patients with colorectal cancer liver metastases (CRLM), however, little study has been done to explore those efficacies using machine learning methods. The present study aimed to evaluate the clinical efficacy of CT texture analysis using machine learning methods as a predictive marker of systemic chemotherapy in patients with CRLM. Methods: Sixty-four patients with CRLM who received first-line chemotherapy were included. Texture analysis was performed on 92 features (First Order Statistics, Gray Level Cooccurrence Matrix, Gray Level Run Length Matrix, Gray Level Size Zone Matrix, Neighbouring Gray Tone Difference Matrix and Gray Level Dependence Matrix) using CT within 1 month before treatment. We evaluated the association between those features and chemotherapeutic response by RECIST (CR+PR vs. SD+PD+NE). We performed eXtreme gradient boost (XGBoost) as a machine learning method to predict the chemotherapeutic response and used the receiver operating characteristic curves to evaluate this prediction model. Results: Main characteristics were the following: male/female = 36/28; median age = 63.5. Patients were treated with oxaliplatin-based chemotherapy (80% of patients), bevacizumab (77%) and anti-EGFR antibody (23%). Thirty-nine patients had confirmed responders, for an overall response rate of 61%, whereas 25 patients (39%) were classified as non-responders (CR: PR: SD: PD: NE = 0: 39: 20: 4: 1). The area under curve of this prediction model was 0.771. Conclusions: We confirmed that CT texture analysis using machine learning for CRLM was feasible. Further analyses are ongoing.