Artificial intelligence for the prediction of synchronous and metachronous liver metastasis in colorectal cancer patients: a systematic review and meta-analysis

Comprehensive Summary

This systematic review evaluated whether artificial intelligence (AI) models can predict synchronous (SLM) or metachronous liver metastasis (MLM) in colorectal cancer, a major contributor to treatment failure and mortality. Using a comprehensive literature search with keywords including AI, deep learning, radiomics, colorectal neoplasm, CT, MRI, and PET-CT, studies were screened under a PICO framework. In total, 21 qualitative studies were identified, with 18 included in a quantitative meta-analysis. All studies were retrospective, most conducted in China, and patient characteristics, tumor stages, and imaging methods varied. Most studies used radiomics, which extracts high-dimensional imaging features (ranging from 107 to 2,632 per case) from segmented regions of interest (ROI). Machine learning methods included random forest, artificial neural networks, genetic algorithms, and ResNet architectures, alongside clinical variables such as TNM stage, extramural vascular invasion (EMVI), and carcinoembryonic antigen (CEA). Quality assessment using QUADAS-2 showed low risk of bias, though the mean Radiomics Quality Score (RQS) was 16.4/36. Meta-analysis using a bivariate random-effects model showed strong performance (sROC AUC = 0.843), with positive predictive value (PPV) of 0.72 and negative predictive value (NPV) of 0.87. Radioclinical models performed better than clinical-only models, and SLM prediction showed higher sensitivity than MLM. There was no significant difference between MRI- and CT-based AI models, suggesting CT-based approaches may improve access and reduce costs without sacrificing accuracy. However, limitations include retrospective designs, lack of standardization, and no cost-effectiveness analysis. Future research should focus on prospective studies, external validation, and improved reproducibility before clinical implementation.

Outcomes and Implications

This research shows that AI-based feature extraction can better capture tumor biology and heterogeneity compared to visual assessment alone. By using radiomics to analyze imaging data, clinicians can gain deeper insight into tumor behavior, supporting more informed and precise clinical decisions. Stronger predictive performance, especially for positive cases, can increase confidence in diagnostic radiology and improve identification of patients at high risk for synchronous (SLM) or metachronous liver metastasis (MLM). Earlier and more reliable prediction of colorectal cancer liver metastasis (CRLM) allows providers to adjust treatment strategies sooner and select more appropriate management options. Overall, integrating AI models into clinical workflows may enhance personalized care, support earlier intervention, and contribute to improved patient outcomes and reduced mortality through more effective risk stratification and screening.