Coupling radiomics analysis of CT image with diversification of tumor ecosystem: A new insight to overall survival in stage I−III colorectal cancer

This study aimed to establish a method to predict the overall survival (OS) of patients with stage I-III colorectal cancer (CRC) through coupling radiomics analysis of CT images with the measurement of tumor ecosystem diversification.We retrospectively identified 161 consecutive patients with stage I-III CRC who had underwent radical resection as a training cohort. A total of 248 patients were recruited for temporary independent validation as external validation cohort 1, with 103 patients from an external institute as the external validation cohort 2. CT image features to describe tumor spatial heterogeneity leveraging the measurement of diversification of tumor ecosystem, were extracted to build a marker, termed the EcoRad signature. Multivariate Cox regression was used to assess the EcoRad signature, with a prediction model constructed to demonstrate its incremental value to the traditional staging system for OS prediction.The EcoRad signature was significantly associated with OS in the training cohort [hazard ratio (HR)=6.670; 95% confidence interval (95% CI): 3.433-12.956; P<0.001), external validation cohort 1 (HR=2.866; 95% CI: 1.646-4.990; P<0.001) and external validation cohort 2 (HR=3.342; 95% CI: 1.289-8.663; P=0.002). Incorporating the EcoRad signature into the prediction model presented a higher prediction ability (P<0.001) with respect to the C-index (0.813, 95% CI: 0.804-0.822 in the training cohort; 0.758, 95% CI: 0.751-0.765 in the external validation cohort 1; and 0.746, 95% CI: 0.722-0.770 in external validation cohort 2), compared with the reference model that only incorporated tumor, node, metastasis (TNM) system, as well as a better calibration, improved reclassification and superior clinical usefulness.This study establishes a method to measure the spatial heterogeneity of CRC through coupling radiomics analysis with measurement of diversification of the tumor ecosystem, and suggests that this approach could effectively predict OS and could be used as a supplement for risk stratification among stage I-III CRC patients.