Predicting the net administered activity in ⁹⁰Y-radioembolization patients from post-procedure ⁹⁰Y-SPECT/CT

Background: The calculation of the net administered activity (A_admin) in patients undergoing ⁹⁰Y-radioembolization is essential for dosimetry and radiation safety, yet current methods for measuring residual ⁹⁰Y activity are often associated with high uncertainty. Therefore, an accurate, robust, and clinically viable method for the determination of A_admin across approved ⁹⁰Y microsphere devices is desirable. Purpose: We report on a novel method to determine A_admin by leveraging the quantitative capabilities of SPECT/CT to measure ⁹⁰Y-emission in vivo from patients following ⁹⁰Y-radioembolization with glass or resin microspheres. Methods: ⁹⁰Y-SPECT/CT attenuation-corrected count data from 147 sequential ⁹⁰Y-radioembolization patients was used for this analysis. A_admin was calculated as part of routine clinical practice via the exposure rate differences between the initial ⁹⁰Y-vial and the ⁹⁰Y-residual jar. This served as our gold standard measure of A_admin. Patient data for each microsphere device were separated into training and testing cohorts to first develop regression models and then to independently assess model performance. The training cohorts were divided into four groups: first, based on the microsphere device (glass or resin), and second, based on the SPECT volume used to calculate counts (the full SPECT field of view (FOV) or liver only (VOI_liver)). Univariate linear regression models were generated for each group to predict A_admin based on ⁹⁰Y-SPECT data from the training cohorts. Leave-one-out cross validation was implemented to estimate variability in model parameters. To assess performance, linear models derived from the training cohort were applied to ⁹⁰Y-SPECT data from the testing cohort. A comparison of the models between microspheres devices was also performed. Results: Linear models derived from the glass and resin training cohorts demonstrated a strong, positive correlation between ⁹⁰Y-SPECT image counts and A_admin for VOI_liver and FOV with R² > 0.98 in all cases. In the glass training cohort, model accuracy (100%—absolute mean prediction error) and precision (95% prediction intervals of mean prediction error) were 99.0% and 15.4% for the VOI_liver and 99.7% and 17.5% for the FOV models, respectively. In the resin training cohort, the corresponding values were 98.6% and 16.7% for VOI_liver and > 99.9% and 11.4% for the FOV models, respectively. The application of these linear models to ⁹⁰Y-SPECT data from the testing cohort showed A_admin prediction errors to have high accuracy and precision for both microsphere devices. For the glass testing cohort, accuracy (precision) was 96.9% (19.6%) and 98.8% (21.1%) for the VOI_liver and FOV models, respectively. The corresponding values for the resin training cohort were 97.3% (26.2%) and 98.5% (25.7%) for the VOI_liver and FOV models, respectively. The slope of the linear models between the two microsphere devices was observed to be significantly different with resin microspheres generating 48%–49% more SPECT counts for equivalent ⁹⁰Y activity based on each device manufacturer's activity calibration process. Conclusion: ⁹⁰Y-SPECT image counts can reliably predict (accuracy > 95% and precision < 18%) A_admin after ⁹⁰Y-radioembolization, with performance characteristics essentially equivalent for both glass and resin microspheres. There is a clear indication that activity calibrations are fundamentally different between the two microsphere devices.