Automatic Prediction of Metastasis-Free Survival (MFS) Using Prostate-Specific Membrane Antigen (PSMA) PET for Oligometastatic Castration-Sensitive Prostate Cancer (omCSPC)

Abstract

Recent data have shown the promise of metastasis-directed therapy (MDT) to improve outcomes in omCSPC patients. Response biomarkers are still needed to identify patients at an early time point who will not respond to the treatment. We have shown that PSMA PET-CT SUVmax changes over time may be a useful response biomarker that correlates with MFS, an important endpoint in these omCSPC patients. This study investigated radiomic imaging biomarkers derived from PSMA-PET acquired pre- and post-MDT for MFS prediction, which may provide better features to discriminate response in the future to improve the outcome of these patients. We accrued an international multi-institutional cohort of omCSPC patients treated with stereotactic ablative radiation therapy (SABR) MDT. The cohort includes 32 patients from Institution 1 (USA) and 38 patients from Institution 2 (Europe). Both pre- and 6-month post-treatment PSMA-PET/CT were acquired. Combat was used for data harmonization in the image domain to minimize imaging variations across institutions. We defined the GTV volume as zone 1 and a 3-5 mm expansion ring area outside the GTV as zone 2 for radiomics analysis. 874 radiomics features (214 original and 660 wavelet filtered features) were extracted from both zones using open-source software and used together for MFS prediction. Function Chi2 was used to select the most significant five features. Several machine learning models (Random Forest, Logistic regression, Support Vector Machine, Naïve Bayesian) were implemented to predict MFS. The models were tested using both a leave-one-out strategy and cross-validation across the two institutions. In the leave-one-out biomarker using 70 patients, random forest achieved the best accuracy, with MFS predicted correctly for 56 (80% of 70) patients. The five radiomic features identified based on their ability to predict MFS included Entropy, Skewness, and Compactness from zone 1, Skewness, and Mesh volume from zone 2. In the cross-institution tests, random forest predicted MFS correctly for 24 (75% of 32) patients when being trained using 38 Institution 2 patients and validated against 32 Institution 1 patients. Vice versa, the model predicted MFS correctly for 28 (74% of 38) patients when being trained using Institution 1 patients and validation using Institution 2 patients. The five features identified for prediction included Entropy and Skewness from both zones and Flatness from zone 1. Our study demonstrated the promise of using pre- and post-MDT PSMA-PET-based imaging radiomic biomarkers for MFS prediction for omCSPC patients. Imaging biomarkers predictive of MFS were identified in both GTV and the ring area outside GTV. Over 74% prediction accuracy was achieved in the cross-institution validation test. The model provides a valuable tool for prognosis prediction early following MDT, which opens up a unique opportunity for monitoring or treatment interventions for patients identified with poor prognoses to improve outcomes.