Abstract / Summary
It has been shown that absorbed dose (AD) to whole-body tumor volume (WBTV) from [ 177 Lu]Lu-PSMA-617 radiopharmaceutical therapy (RPT) correlates with treatment response. Pre-therapy estimation of AD to WBTV could therefore aid clinicians with patient selection and dosimetry-guided treatment planning. The goal of this work was to predict post-cycle 1 [ 177 Lu]Lu-PSMA-617 RPT AD to WBTV from standard-of-care pre-therapy [ 68 Ga]Ga-PSMA-11 PET metrics and blood biomarkers. Baseline [ 68 Ga]Ga-PSMA-11 PET/CT scans were obtained prior to initiating [ 177 Lu]Lu-PSMA-617 RPT, and 3–4 SPECT/CT scans were performed after cycle 1. Whole-body tumor volume (WBTV) for each patient was segmented using a commercially available, semi-automated tool and was verified by a radiologist. Additionally, up to 6 index tumors per patient were segmented, while kidneys and salivary glands were segmented via deep-learning tools. Six PET metrics and 15 biomarkers were evaluated as predictors. We developed univariable and multivariable linear regression models using 3 different features sets (biomarker-only, PET-only, and PET+biomarker), as well as random forest regression machine-learning models incorporating all features, to predict cycle 1 AD to WBTVs, index tumors, kidneys, and salivary glands. Feature selection for linear regression models was performed using least absolute shrinkage and selection operator (LASSO) with a stability selection threshold of 0.7. Model performance was assessed using leave-one-patient-out cross-validation with root mean squared error (RMSE) and median absolute percent error (median APE). Sensitivity analyses were performed to assess the performance of random forest classification and logistic regression models for predicting binary prostate-specific antigen (PSA) response, defined as AD to WBTV above previously identified thresholds. Thirty-one patients were included in this analysis. SUV Mean was the only feature selected from the PET-only and PET+biomarker feature sets for WBTV, and the resulting linear regression model achieved cross-validated median APE of 42% and RMSE of 0.69 Gy/GBq for predicting AD. The random forest model offered improved median APE (27%) but slightly higher RMSE (0.76 Gy/GBq). Linear regression models achieved similar, but lower, RMSE and median APE compared to random forest models for index tumor (RMSE: 2.51 Gy/GBq, median APE: 56%), kidney (0.20 Gy/GBq, median APE: 32%), and salivary gland (0.16 Gy/GBq, median APE: 28%). Blood biomarkers little no predictive value beyond PET-derived metrics for linear regression models. Sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) for predicting binary PSA response were high for both random forest classification and logistic regression models. Cycle 1 [ 177 Lu]Lu-PSMA-617 AD to WBTVs, index tumors, kidneys, and salivary glands can be predicted using pre-therapy [ 68 Ga]Ga-PSMA-11 PET metrics. Across all structures, random forest and linear regression models provided comparable prediction performance. However, to be implemented clinically these findings must be verified using larger and external data sets.