576 - Radiomics-Based Imaging Biomarkers for Predicting Response to TARE: A Systematic Review and Meta-Analysis

Purpose: Radiomics offers a noninvasive means to predict outcomes after transarterial radioembolization (TARE), yet its reproducibility and performance remain unclear. This systematic review and meta-analysis assesses predictive accuracy, imaging modalities, and modeling strategies, providing tumor type and modality pooled metrics.

Materials and Methods: Following PRISMA guidelines, PubMed, Embase, Scopus, and Web of Science were searched through April 2024. Studies using radiomics to predict treatment response or survival after TARE were included. Quality was assessed via QUADAS-2 and Radiomics Quality Score (RQS). Data on tumor type, imaging modality, model type, validation, and performance metrics were extracted. Meta-analysis of 11 externally validated studies was performed using a random-effects model to estimate pooled sensitivity, specificity, and AUC, with subgroup analyses by imaging type and tumor origin.

Results: Out of 608 screened articles, 14 studies (1,423 patients) were included. Tumors studied included HCC (n=9), iCC (n=2), and mCRC (n=3). Imaging sources included CT (n=3), MRI (n=6), and PET (n=2), with some using hybrid or multiparametric inputs. Meta-analysis of 11 studies with quantitative outcomes revealed a pooled AUC of 0.82 (95% CI: 0.78–0.86), sensitivity of 79% (95% CI: 72–85%), and specificity of 75% (95% CI: 67–82%), with moderate heterogeneity (I² = 52%). Subgroup analysis demonstrated superior performance in PET-derived models (AUC = 0.87) compared to CT (0.81) and MRI (0.78). Studies integrating radiomics with clinical features (e.g., AFP, tumor burden) achieved higher AUCs (0.85) than radiomics-only models (0.76). Meta-regression showed that studies using external validation and those with higher RQS ( >12) had significantly better model robustness (p < 0.01). Temporal resolution (e.g., baseline vs. follow-up imaging), feature reduction strategies (e.g., LASSO vs. PCA), and inclusion of shape descriptors were key moderators of model performance. Notably, studies using 3D radiomics and those focusing on texture heterogeneity features (e.g., entropy, GLCM contrast) had better predictive metrics across tumor types. Predicting complete or partial response was more accurate than predicting stable/progressive disease.

Conclusion: Radiomics models effectively predict treatment response after TARE, particularly when combined with clinical data and PET imaging. However, methodological variability and reporting gaps underscore the need for standardization and prospective validation.