3199 - Artificial Intelligence Model Using PSMA PET Imaging of Primary Prostate Cancer for Analyzing Metastases and Post-Treatment Progression

J. E. Juarez Casillas¹, C. M. Benitez², K. Rzechowski³, K. Sjöstrand³, A. Anand³, I. Sonni⁴, G. Berenji⁵, S. S. Duriseti⁶, M. Rettig⁷, N. Kane⁸, and N. G. Nickols⁹; ¹Department of Radiation Oncology, University of California Los Angeles, Los Angeles, CA, ²Department of Radiation Oncology, University of California, Los Angeles, Los Angles, CA, ³Exini Diagnostics AB, Lund, Sweden, ⁴Department of Radiological Sciences, Los Angeles, CA, ⁵VA Greater Los Angeles Healthcare System, Los Angeles, CA, ⁶Washington University School of Medicine, Department of Radiation Oncology, St. Louis, MO, ⁷Department of Medical Oncology, University of California, Los Angeles, Los Angeles, CA, ⁸UCLA David Geffen School of Medicine, Los Angeles, CA, ⁹University of California Los Angeles, Department of Radiation Oncology, Los Angeles, CA

Purpose/Objective(s): Piflufolastat F-18 DCFPyL (PyL) is a PSMA targeted imaging agent that provides whole-body staging of prostate cancer. Image analysis of the primary tumor using machine learning algorithms might offer additional insight into disease biology. In this proof-of-concept study, we developed convolutional neural network (CNN) models using inputs from the entire prostate on PyL images along with auto-segmented hotspots within the primary prostate tumor to predict the presence of synchronous metastases in patients with newly diagnosed prostate cancer. Materials/

Methods: Veterans with de novo prostate cancer imaged with PyL PSMA PET/CT for initial staging were included in this retrospective analysis. The PyL scans were analyzed using aPROMISE, which automatically segments, localizes, and quantifies disease on PSMA PET/CT. The automated CT segmentations of the prostate were used to map the PyL PET uptake of the prostate. Both the entire prostate, as well as the auto-segmented aPROMISE-defined hotspots were used as inputs for the CNN. The dataset was randomly split into training, validation and test sets. The area under ROC curve (AUC) was computed to determine the performance of the model in predicting the presence of metastases and the test predictions were compared with ground truth (M1). Training was repeated 50 times and the best performing experiment was identified. Prediction scores from UCSF-CAPRA and UCLA PSMA risk calculator were used as comparators. An additional dataset consisting of 20 patients with either early metastatic progression (= 3 years) or long-term post treatment response (> 4 years) without any progression was used to determine whether the model could prognosticate metastatic progression.
Results: 90 Veterans with de novo localized (n=47) and metastatic (n=43) prostate cancer were included in this analysis. The CNN model had a median AUC of 0.72 for prediction of metastatic disease (ICR 0.64 and 0.8). Addition of clinico-pathologic and measurable imaging data to the CNN data in a fused model improved the AUC to a median of 0.82. In the separate dataset of patients with either early metastatic progression (n=10) or long-term post treatment response (n=13), the CNN model had an AUC of 0.72, while the multimodal model achieved an AUC of 0.84 for predicting metastatic progression.
Conclusion: The CNN model using PyL imaging demonstrates that synchronous metastases can be predicted from intraprostatic PyL uptake patterns alone with a predictive accuracy comparable to published prediction models based on clinico-pathologic features, and further enhanced with the addition of clinico-pathologic and measurable imaging data. This study suggests that PyL CNN-based models could be developed to prognosticate metastatic progression, and further research is needed.