Synolitic Graph Neural Networks for MRI-Derived Radiomic-Based Prediction of Prostate Cancer Progression on Active...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Prostate cancer (PCa) is a common malignancy in men.
• Active surveillance (AS) is standard for low-risk PCa, but predicting progression is challenging.
• MRI-derived radiomics aids risk stratification, yet conventional methods overlook feature interdependencies.

Purpose of the Study:

• To evaluate Synolitic Graph Neural Networks (SGNNs) for predicting PCa progression in AS patients.
• To assess SGNNs' ability to model complex inter-feature relationships in radiomic data.
• To compare SGNN performance against traditional machine learning models.

Main Methods:

• Utilized a cohort of 343 AS patients (73 progressors, 270 non-progressors).
• Extracted 72 radiomic features from baseline 3T MRI and 3 clinical variables.
• Developed an SGNN pipeline transforming features into patient-specific graphs, evaluated GCN and GATv2 architectures, and compared against Gradient Boosting, SVM, Random Forest, and logistic regression.

Main Results:

• SGNN with GATv2 and confidence-based sparsification achieved the highest ROC-AUC (0.699 ± 0.044).
• This represents an absolute improvement of 0.065 over the best conventional method (Gradient Boosting, ROC-AUC = 0.634 ± 0.080).
• Incorporating topological node features consistently improved performance by 3-5%.

Conclusions:

• SGNN framework demonstrates potential as a complementary paradigm for radiomic data analysis in oncology.
• The approach effectively models inter-feature relationships, outperforming conventional methods in predicting PCa progression.
• Further validation in external cohorts and incorporation of longitudinal data are warranted.