Author: Clint Bahler, Ruchika Reddy Chimmula, Harrison Louis Love, Oluwaseyi Oderinde, Courtney Yong 👨🔬
Affiliation: Purdue University, Department of Urology, Indiana University School of Medicine, Advanced Molecular Imaging in Radiotherapy (AdMIRe) Research Laboratory, School of Health Sciences, Purdue University 🌍
Purpose: Prostate cancer (PCa) is a common malignancy in men, and predicting biochemical recurrence (BCR) is crucial for guiding treatment decisions. Integrating multimodal data, including clinical, imaging (PET, MRI), and genomic tests like Decipher, can enhance predictive models. Fusion techniques, particularly advanced methods such as intermediate fusion with feature selection, have the potential to improve prediction accuracy and support precise clinical decision-making in PCa management.
Methods: Data from 108 PCa patients who underwent radical prostatectomy (RP) at Indiana University were analyzed. The cohort had a median age of 61.5 years (range: 48–77), with median preoperative PSA levels of 7.15 ng/ml (range: 2.7–32). Key clinical indicators included PSA density, fraction of positive biopsy cores, Gleason grades, and ISUP biopsy grades, alongside 68Ga-PSMA-11 PET, MRI, and Decipher genomic test results. Early fusion combined all modalities at the input level, while intermediate fusion was implemented in two ways: (1) feature selection to emphasize significant predictors, and (2) weighted modalities based on feature importance. Machine learning models, including Random Forest and XGBoost, were applied to evaluate performance.
Results: Intermediate fusion with feature selection consistently outperformed other methods in leveraging multimodal data for prediction. The XGBoost model with feature selection achieved an accuracy of 87.5% and an AUC of 0.93, outperforming intermediate fusion with weighted modalities (accuracy: 75%, AUC: 0.81) and early fusion XGBoost (accuracy: 85%, AUC: 0.91). For Random Forest, intermediate fusion with weighted modalities achieved an accuracy of 75% and an AUC of 0.84, while early fusion performed better. These findings highlight the superior performance of feature selection in intermediate fusion.
Conclusion: Intermediate fusion with feature selection (XGBoost) is a superior technique for predicting BCR in PCa patients, combining clinical, PET, MRI, and genomic data to improve diagnostic accuracy. This approach can guide personalized treatment strategies, optimizing outcomes in PCa management.