- 2025-11-28
- JMIR medical informatics 13
- Yuan Sun
- Bo Li
- Chuanlan Ju
- Liming Hu
- Huiyi Sun
- Jing An
- Tae-Hun Kim
- Zhijun Bu
- Zeyang Shi
- Jianping Liu
- Zhaolan Liu
Study Design
- Type
- Meta-Analysis
- Sample size
- n = 4,600
- Population
- 4600 patients with CRC
- Methods
- Systematic search of 4 databases (Embase, PubMed, the Cochrane Library, and Web of Science) up to January 1, 2025; included studies developing or validating radiomics-based machine learning models for predicting CRC recurrence using CT or MRI; used a bivariate mixed-effects model for meta-analysis
Background
Predicting colorectal cancer (CRC) recurrence risk remains a challenge in clinical practice. Owing to the widespread use of radiomics in CRC diagnosis and treatment, some researchers recently explored the effectiveness of radiomics-based models in forecasting CRC recurrence risk. Nonetheless, the lack of systematic evidence of the efficacy of such models has hampered their clinical adoption.Objective
This study aimed to explore the value of radiomics in predicting CRC recurrence, providing a scholarly rationale for developing more specific interventions.Methods
Overall, 4 databases (Embase, PubMed, the Cochrane Library, and Web of Science) were searched for relevant articles from inception to January 1, 2025. We included studies that developed or validated radiomics-based machine learning models for predicting CRC recurrence using computed tomography or magnetic resonance imaging and provided discriminative performance metrics (c-index). Nonoriginal articles, studies that did not develop a model, and those lacking clear outcome measures were excluded from the study. The quality of the included original studies was assessed using the Radiomics Quality Score. A bivariate mixed-effects model was used to conduct a meta-analysis in which the c-index values with 95% CI were pooled. For the meta-analysis, subgroup analyses were conducted separately on the validation and training sets.Results
This meta-analysis included 17 original studies involving 4600 patients with CRC. The quality of the identified studies was low (mean Radiomics Quality Score 13.23/36, SD 2.56), with limitations in prospective design and biological validation. In the validation set, the c-index values based on clinical features, radiomics features, and radiomics features combined with clinical features were 0.73 (95% CI 0.68-0.79), 0.80 (95% CI 0.75-0.85), and 0.83 (95% CI 0.79-0.87), respectively. In the internal validation set, the c-index values based on clinical features, radiomics features, and radiomics features+clinical features were 0.70 (95% CI 0.61-0.79), 0.83 (95% CI 0.78-0.88), and 0.83 (95% CI 0.78-0.88), respectively. Finally, in the external validation set, the c-index values based on clinical features, radiomics features, and radiomics features combined with clinical features were 0.76 (95% CI 0.70-0.83), 0.75 (95% CI 0.66-0.83), and 0.83 (95% CI 0.78-0.88), respectively.Conclusions
Radiomics-based machine learning models, especially those integrating radiomics and clinical features, showed promising predictive performance for CRC recurrence risk. However, this study has several limitations, such as moderate study quality, limited sample size, and high heterogeneity in modeling approaches. These findings suggest the potential clinical value of integrated models in risk stratification and their potential to enhance personalized treatment, though further high-quality prospective studies are warranted.