The liver is the organ most often affected by distant metastases of CRC, and many patients develop metastases despite successful resection. In this study, we evaluated a radiomics model that combined CT features and clinical data in the hope that it could be used as a non-invasive tool for the preoperative prediction of liver metastases after surgery for CRC. Our model showed good performance in both the training and validation sets (with respective AUCs of 0.952 and 0.761), indicating that the integration of radiomics features and clinical factors was satisfactory in all cohorts. In this model, tools that can confidently detect colorectal malignant metastases preoperatively may be valuable for patient management. Therefore, the imaging-clinicopathomics model has better performance and stability while processing different types of samples at different thresholds, while maintaining good sensitivity and accuracy [17].

Many researchers have investigated the relationship between imaging characteristics and tumor metastases. Studies by Lee et al and Taghavi et al have also shown that CT radiomics features and clinical features were effective in predicting liver metastases after surgery for CRC (with respective AUCs of 0.747 and 0.86). The study by Becker et al, which included 165 patients, found that whole-liver CT texture could predict patients at risk of developing liver metastases within ≤ 6 months (AUC 0.74) but not at later stages [18]. These findings suggest that metachronous metastases can develop in response to microenvironmental changes in patients with no apparent liver involvement at the time of diagnosis of CRC. Our studies have focused predominantly on CRC features rather than liver features, and some studies have shown that heterogeneity in primary CRC tumors is associated with the aggressiveness of liver metastases from CRC and can predict their likelihood. A combination of clinical features and radiomics features is useful for understanding the heterogeneity of colorectal tumors and their surrounding microenvironment. In view of the better diagnosis of colorectal cancer metastases, we emphasize the maximum stability and generalization ability, and more accurately diagnose the risk of missing potentially life-threatening diseases [17]. Among the various types of cancer, there is ample evidence that the metabolic state of tumors during metastasis is constantly adapting to the microenvironment [19, 20]. In studies by Liu et al and Shu et al, radiomics features obtained from magnetic resonance images of primary rectal tumors predicted synchronous liver metastases [21, 22].

Metachronous liver metastases significantly affect the prognosis of patients with CRC after radical colorectal resection. Previous studies have identified age, TNM stage, and tumor markers (CEA and CA19-9) to be clinical risk factors for colorectal metastasis [23,24,25,26,27,28]. Hao et al used clinical data to establish a novel nomogram, the AUC for which was 0.786, and the verification AUC was 0.784, which had a good differential effect on metachronous liver metastasis. Although these factors have good predictive performance, they cannot accurately determine the clinical stage or size of the negative tumor and the obsolescence of nomograms in radiomics research [29]. Other factors may also predispose to metachronous liver metastasis in CRC, including changes in the tissue microenvironment and tumor heterogeneity [30]. In our study, we used logistic regression and random forest classification, very common and widely studied machine learning models. For the prediction of liver metastases, the model trained by logistic regression showed the importance of adding imaging features to the clinical features compared with the ones only consisting of clinical features.

In this study, LASSO analysis (L1 regularization) and t-tests identified 14 signature radiomics features out of 1051 radiomics features, including three original image features and eleven wavelet image features. Tumor size is an important predictor of risk stratification in abdominal CT features. Maximum tumor diameter and shape were positively correlated with the risk of metastasis. First-order statistical characteristics reflect the degree of dispersion of gray values in the ROI. The higher the degree of malignancy of the tumor, the more the dispersion. Research by Cambria et al has shown that the potential for metastasis is closely related to the intratumoral environment [31]. Higher-order texture features reflect the uneven distribution of image texture in comprehensive information such as space and distance. This suggests that tumors with high malignant potential have complex internal structures and are more prone to heterogeneity.

This study has some strengths. First, unlike previous studies, it obtained clinical data that correlated with liver function and recurrence of CRC. Second, segmentation was based on the entire three-dimensional volume of the tumor, and radiomics features were extracted using a machine learning algorithm to maximize the potential information underlying the images and identify features with the highest predictive value. Random forests can process high-latitude data without feature selection, are strongly resistant to overfitting, and can detect interaction between features.

There are also some limitations to our study. First, it had a retrospective single-center design. A prospective multicenter study is required in the future. Second, as a retrospective study with a limited data set, selection bias and the presence of unknown confounders were possible. Although we tried to minimize selection bias by using rigorous inclusion criteria, we could not rule out unknown confounders that could have affected the results. Third, in our study, the sample was predominantly from the same region, and this imbalance in demographic and geographic distribution may have limited the model’s understanding of population characteristics such as different regions and ethnicities. To mitigate the impact of this potential bias, we will employ a broader sample selection strategy in future studies to ensure that different regions and population groups are adequately represented.

In conclusion, this study found that a model that included CT radiomic-based features and clinical features using the RL algorithm was better able to predict liver metastases in patients with CRC than a radiomics model or clinical model alone.