Predicting PSA50 response to [Formula: see text]Lu-PSMA therapy using machine learning and automated total tumor volume.
Authors
Affiliations (6)
Affiliations (6)
- CREATIS; CNRS UMR 5220; INSERM U 1044; Université de Lyon; INSA-Lyon; Université Lyon 1, Lyon, France. [email protected].
- Centre de lutte contre le cancer Léon Bérard, Lyon, France. [email protected].
- Siemens Healthcare, Courbevoie, France. [email protected].
- Centre de lutte contre le cancer Léon Bérard, Lyon, France.
- Centre Hospitalier de Saint Denis, La Réunion, France.
- CREATIS; CNRS UMR 5220; INSERM U 1044; Université de Lyon; INSA-Lyon; Université Lyon 1, Lyon, France.
Abstract
The <sup>177</sup>Lu-PSMA therapy is an established treatment for metastatic castration-resistant prostate cancer (mCRPC), targeting the prostate-specific membrane antigen (PSMA). Despite well-established correlations between <sup>68</sup>Ga-PSMA PET/CT imaging and outcome, predicting individual patient responses remains a significant challenge. This study introduces an automated method for computing the total tumor volume (TTV) from <sup>68</sup>Ga-PSMA PET/CT imaging and develops predictive models to assess patient biological response via the PSA50 criterion. A retrospective analysis was conducted on a real-world data cohort of 139 mCRPC patients treated in our institution. TTV was automatically extracted from PET/CT images and correlated with treatment response, defined by PSA50 criteria. Machine learning models, including Logictic Regression with L1 (LASSO) and Support Vector Machine (SVM), were developed to predict PSA50 response using imaging and clinical features. The best-performing models achieved F1-scores of 0.68 and 0.67, comparable to existing nomograms. Correlation analysis identified TTV-derived features and time since diagnosis as significant predictors of response. The proposed workflow offers an automated and reproducible approach to predicting treatment response in <sup>177</sup>Lu-PSMA therapy. Limitations remain for lesion segmentation within physiological regions.