AI-Based MRI Radiomics for Preoperative Detection of Tumor Deposits in Rectal Cancer: A Bivariate Diagnostic Accuracy Meta-analysis

Background: In rectal cancer (RC), tumor deposits (TDs) are clinically important as they are associated with a negative impact on prognosis and affect risk stratification as well as management plans. MRI radiomics combined with artificial intelligence (AI) has been explored for non-invasive preoperative TDs detection. However, results vary across studies; therefore, we performed a systematic review and meta-analysis to pool their diagnostic test accuracy (DTA). Methods: This systematic review and DTA meta-analysis were conducted in accordance with the Cochrane Handbook for Systematic Reviews of Diagnostic Test Accuracy and reported in adherence to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy Studies (PRISMA-DTA). Databases were screened for DTA studies of MRI-based radiomics predicting TDs in adult RC patients with histopathology as reference. Results: Nine studies (2589 patients) were identified. Across included studies, the bivariate meta-analysis showed pooled sensitivity of 0.765 (95% CI 0.714 - 0.809) and specificity of 0.790 (95% CI 0.750 - 0.825), with overall AUC of 0.845 and very low heterogeneity (I2 = 7.3%). In validation datasets, diagnostic performance remained consistent (AUC 0.806), with pooled sensitivity 0.751 (95% CI 0.681 - 0.809) and specificity 0.751 (95% CI 0.681 - 0.809). For MRI modality, T2 weighted image (T2WI) showed higher sensitivity (0.745) than T2WI plus diffusion weighted image (DWI) (0.697), while T2WI+DWI had higher specificity (0.809 vs 0.737) and AUC (0.829 vs 0.796), with no significant subgroup differences (p≥0.192). Models combining radiomics with clinical/radiological variables achieved significantly higher sensitivity (0.842 vs 0.657, p<0.001) and higher AUC (0.827 vs 0.752), with comparable specificity. Conclusion: MRI-based radiomics demonstrates good accuracy for preoperative detection of tumor deposits in RC. The integration of radiomics with clinical or imaging features increases the accuracy of identification. Amongst the MRI modalities, models that utilize the combination of T2WI plus DWI are more accurate than those using T2WI or DWI only.