Abstract
BACKGROUND: Interpreting [(18)F]PSMA-1007 PET/CT scans can be challenging due to the occurrence of unspecific uptake in lymph nodes and bones. Machine learning has proven its suitability to use features to model complex relations leading to accurate diagnosis. We aimed to investigate the impact of contextual information on machine learning performance in identifying lymph node and bone metastases in prostate cancer patients using [(18)F]PSMA-1007 PET. A Random Forest Classifier (RFC) and Extreme Gradient Boosting (XGBoost) were trained across two feature sets to classify hotspots into malignant or non-malignant. The first set incorporated hotspot-specific features, such as SUVmax, anatomic location and tissue type of the location (lymph node/bone). The second set was the first set combined with context-level features, such as SUVmax of nearby hotspots and the number of hotspots. RESULTS: We retrospectively included 103 patients who underwent clinically indicated [(18)F]PSMA-1007 PET/CT, in whom hotspots were observed in lymph nodes (n = 256) and bone structures (n = 267). The context-enhanced model outperformed the hotspot-specific model in Area Under The Curve (AUC) and Youden Index for both RFC and XGBoost. The context-enhanced RFC performed superior in AUC for bone (0.92, p < 0.001) and lymph node hotspots (0.95, p < 0.001). The performance increase after adding contextual information was stronger for bone hotspots compared to lymph node hotspots in terms of AUC (0.06 vs. 0.03, p < 0.001) and Youden Index (0.15 vs. 0.07, p < 0.001). CONCLUSION: We successfully developed models to accurately identify lymph node and bone metastases. We underscored the potential of leveraging contextual information in machine learning methods to improve the identification of lymph node and bone metastases.