Fig. 1.

Study design. Overview of the analysis framework. Feature extraction on the abdominal CT, 50 × 50 pixel ROIs, was done by utilizing a pre-trained convolutional neural network. We preprocessed them based on the significance of association with 5-year liver metastasis (5YLM) rate by performing univariate logistic regression analysis. Principal component analysis (PCA) was done for feature reduction in dimensionality and this generated new sets of feature. We used two machine learning algorithms, such as logistic regression classification (LR) and random forest classification (RFC) to train prediction models for 5YLM and compared the performances of each model. Among the models to predict 5YLM, we used the highest AUC model to perform multivariate logistic regression to association between the image features and 5YLM statistically. Then Kaplan Meier analysis was done by the principal components (PCs) for metachronous liver metastasis free survival and overall survival. We done a correlation analysis between the significant PCs and the clinical variable in Table 1. We also applied the highest AUC model for 5YLM to predict 5-year mortality and observed whether the liver image feature could do a predictive role for 5-year mortality