Leveraging the non-contrast CT component of PET/CT: an AI-driven delta-radiomics approach to monitor treatment response in metastatic breast cancer.

PURPOSE: 18 F-FDG PET/CT is the standard modality for monitoring treatment response in metastatic breast cancer. This study aims to evaluate the predictive value of delta-radiomics derived solely from the low-dose, non-contrast CT component acquired during routine PET/CT imaging-without requiring an additional dedicated CT examination or extra contrast administration-for monitoring response to CDK4/6 inhibitors in de novo metastatic hormone receptor-positive (HR+)/HER2-negative breast cancer.

METHODS

This retrospective study included 33 patients with bone-predominant metastatic breast cancer. Delta radiomic features were extracted from the non-contrast CT component of paired baseline and follow-up 18 F-FDG PET/CT scans. Patients were stratified into Responders (Complete or Partial Response) and Non-Responders (Stable or Progressive Disease) based on standard PERCIST criteria. We developed an integrated machine learning model using logistic regression with elastic net regularization, validated via leave-one-out cross-validation (LOOCV).

RESULTS

The cohort consisted of 25 Responders and 8 Non-Responders. Non-Responders exhibited distinct longitudinal increases in Delta_Pct_shape_Elongation and Delta_Pct_firstorder_90Percentile compared to Responders. The integrated model, combining these features with clinical variables, achieved an Area Under the Curve (AUC) of 0.930, significantly outperforming the baseline clinical-only model (AUC = 0.775). While the default threshold prioritized sensitivity (96.0%) with limited specificity (25.0%), post-hoc threshold optimization maximizing the Youden index demonstrated a highly balanced performance, achieving 88.0% sensitivity and 87.5% specificity.

CONCLUSIONS

Delta radiomics analysis of the routinely acquired non-contrast CT component of PET/CT provides substantial incremental prognostic value over standard clinical variables. This approach demonstrates the potential of utilizing existing low-dose CT data as a cost-effective, supportive biomarker for the early prediction of therapeutic resistance.