A Graph Attention Network Combining Both Radiomics and Clinical Data for Improved and Interpretable Prediction of Lymph Node Invasion in High-Grade Prostate Cancers
M Larose*1, 2, 3, N Touma3, N Raymond4, D Leblanc1, 2, 3, F Rasekh3, B Neveu3, H Hovington3, M Vallières4, F Pouliot3, L Archambault1, 2, 3, (1) Department of physics, Université Laval, Qc, Canada (2) Centre de recherche sur le cancer, Qc, Canada (3) CHU de Québec, QC, Canada (4) Department of computer science, Université de Sherbrooke, QC, Canada
Presentations
SU-H400-IePD-F5-1 (Sunday, 7/10/2022) 4:00 PM - 4:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 5
Purpose: To investigate a graph attention network (GAT) based approach combining both radiomics and clinical data for improved and interpretable prediction of lymph node invasion (LNI) in high-grade prostate cancers (PCa).
Methods: Experiments were conducted using an IRB approved in-house dataset of 170 high-grade PCa (Gleason≥8) each with FDG-PET/CT images acquired prior to prostatectomy for staging purposes. A total of 101 and 94 radiomic markers were extracted from the CT and PET images respectively and ordered by Gini importance. The population graph was built using the k-nearest neighbor graph construction method with a mutual information weighted similarity metric. To ensure learning of a graph with interpretable connections, the features used for graph construction were the clinical features merged with the most important radiomic shape-based CT feature and first-order intensity-based PET feature. As for the node features, we combined clinical data and the top 10% of radiomic markers. The performance of the GAT model was compared to a random forest classifier (RFC) and a support vector machine (SVM). The models were optimized using a stratified nested 5-fold cross-validation with 140 patients and evaluated on a holdout-set of 30 patients. Adjustments for class imbalance were applied during modeling.
Results: On the validation sets, the models reached {AUC=0.68±0.09, bACC=0.65±0.07}, {AUC=0.61±0.05, bACC=0.61±0.06} and {AUC=0.68±0.07, bACC=0.68±0.07} for the GAT, RFC and SVM models respectively. On the holdout-set, the models reached {AUC=0.765, bACC=0.705}, {AUC=0.66, bACC=0.66} and {AUC=0.725, bACC=0.725}. Although SVM achieved better balanced accuracy than GAT, the predictions made by the latter have the advantage of being interpretable through the graph-attention mechanism, which aims to assign greater importance to neighbors of the same class.
Conclusion: Our results suggest that combining imaging and non-imaging information in an intuitive graph-based approach improves the interpretability of LNI predictions in high-grade PCa while maintaining good performance compared to other models.
Funding Support, Disclosures, and Conflict of Interest: Natural Sciences and Engineering Research Council of Canada (NSERC), Fonds de Recherche du Quebec - Nature et technologies (FRQNT)
Keywords
CAD, FDG PET, Feature Selection
Taxonomy
Contact Email