Radiomic clustering using graph network techniques coupled with unbalanced optimal transport

Area of Science:

• Oncology
• Radiology
• Bioinformatics
• Machine Learning

Background:

• High-dimensional radiomic datasets often require subgroup identification for effective machine learning application.
• Limited dataset sizes pose challenges for traditional machine learning methods in cancer research.

Purpose of the Study:

• To develop and validate a novel approach for identifying radiomic subgroups in cancer datasets.
• To investigate the prognostic value and immune cell associations of identified radiophenotypes in head and neck squamous cell carcinoma (HNSCC) and non-small cell lung cancer (NSCLC).

Main Methods:

• A regularized network model with an extended Bayesian information criterion was used to identify sub-networks.
• A graph network-based k-means clustering algorithm with unbalanced optimal transport was developed for sample grouping.
• Survival analysis (Kaplan-Meier) and CIBERSORT analysis were performed on identified subgroups using radiomic features from CT scans and RNA-Seq data.

Main Results:

• The proposed method identified high- and low-risk groups in HNSCC, showing significant differences in progression-free survival (p=0.0202).
• In NSCLC, distinct high- and low-risk groups were identified with significant differences in overall survival (p=0.0007).
• Significant differences in immune cell abundance (neutrophils in HNSCC; resting dendritic cells and activated mast cells in NSCLC) were observed between risk groups.

Conclusions:

• Radiomic characteristics can effectively identify radiophenotypes with distinct prognoses.
• The identified radiophenotypes may be associated with varying tumor-immune interactions, offering potential for personalized cancer treatment strategies.