3305 - Assessing the Predictive Accuracy of Radiogenomic Models Incorporating the Decipher Genomic Classifier and Radiomic Features for Biochemical Recurrence in Prostate Cancer
Presenter(s)
R. Chimmula1, C. Yong2, H. L. Love2, R. Shiradkar3, J. A. Holmes4, C. Bahler2, and O. M. Oderinde1; 1Advanced Molecular Imaging in Radiotherapy (AdMIRe) Research Laboratory, School of Health Sciences, Purdue University, West Lafayette, IN, 2Department of Urology, Indiana University School of Medicine, Indianapolis, IN, 3Department of Biomedical Engineering and Informatics, Indiana University, Indianapolis, IN, 4Department of Radiation Oncology, Indiana University School of Medicine, Indianapolis, IN
Purpose/Objective(s): Prostate cancer (PCa) significantly affects men's quality of life and survival rates. Primary treatments for localized PCa include radical prostatectomy (RP), radiotherapy, and hormone therapy. Despite these interventions, many patients experience biochemical recurrence (BCR), indicated by rising prostate-specific antigen (PSA) levels post-treatment. Accurately predicting BCR is critical for optimizing treatment strategies and advancing personalized care. This study investigates the predictive performance of radiogenomic models that incorporate the Decipher Genomic Classifier (DGC) and radiomic features from PET/CT images to predict BCR in PCa patients. While DCG and PSMA-PET/CT radiomics have individually demonstrated potential in predicting BCR, their combined impact on enhancing prediction accuracy remains largely unexplored.
Materials/Methods: We retrospectively selected 40 PCa patients who underwent RP, with 60% experiencing BCR within three years. The DGC provides a 22-gene expression-based risk score, and radiomic features were extracted from 68Ga-PSMA-11 PET/CT images within the prostate gland. Predictive models were developed using these variables, along with clinical factors such as International Society of Urological Pathology (ISUP) grade, PSA density, and the fraction of positive cores. We employed Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Logistic Regression (LR) algorithms to build the models. Additionally, we explored various fusion techniques for integrating radiogenomic features. Model performance was assessed using accuracy and area under the curve (AUC) metrics.
Results: The radiogenomic model developed with XGBoost using an intermediate fusion approach achieved the best performance, with an AUC and accuracy of 0.78 (95% CI: 64.5% - 99.5%) and 87.5%, respectively. In addition, radiomic-only models achieved model performance of 0.65 (95% CI: 67.6% - 92.4%), 0.72 (95% CI: 70.7% - 94.3%), and 0.67 (95% CI: 67.6% - 92.4%) for XGBoost, RF, and LR, respectively. In comparison, DGC-only models achieved AUCs of 0.48 (95% CI: 61.5% - 88.4), 0.63 (95% CI: 64.6% - 90.4%), and 0.34 (95% CI: 64.6% - 90.4%), respectively for the same algorithms.
Conclusion: This study demonstrates that an XGBoost radiogenomic model with an intermediate fusion technique has significant potential to improve the prediction accuracy of BCR, thereby supporting patient selection for optimized and personalized treatment strategies. Future research will focus on enhancing model performance by increasing the sample size and validating the model with external data.