Background: RhoB plays a crucial role in cellular processes. Its expression, evaluated via immunohistochemistry (IHC), serves as a significant biomarker for predicting survival outcomes in cancer patients, particularly in response to radiotherapy (RT). However, the accuracy and reliability of RhoB expression assessments can be compromised by variability in staining techniques and tumor heterogeneity.
Methods: This study introduces the scalable recurrence graph network (SRGNet) to improve the accuracy and reliability of five-year survival predictions for patients with rectal cancer. SRGNet integrates principles of spatial statistics, nonlinear dynamics, graph theory, and graph convolutional networks (GCNs). Spatial statistics aims to capture the distribution of protein expression, while nonlinear dynamics can model complex temporal changes within tumor biopsies. Graph theory represents relationships between texture recurrences, and GCN architecture enables graph node classification.
Results: SRGNet outperformed 10 pre-trained convolutional neural networks. For biopsy samples with RT, SRGNet achieved an accuracy of 88%, predicting <five-year survival with 67% accuracy and >five-year survival with 100% accuracy. It demonstrated a precision of 100% and an F1 score of 0.80, with an AUC of 0.73. For biopsy samples without RT, SRGNet obtained an accuracy of 91%, predicting survival over five years with perfect precision (100%), an F1 score of 0.86, and an AUC of 0.82.
Conclusions: SRGNet can enhance the biopsy-based prediction of five-year survival outcomes for rectal cancer patients, offering a more precise and individualized assessment. Its integration of artificial intelligence and machine learning technologies addresses the challenges of variability and tumor heterogeneity in RhoB expression evaluation, leading to more reliable prognostic evaluations.