Enhancing PET/CT Radiomics Robustness Through Graph Signal Processing

Area of Science:

• Medical Imaging
• Radiomics
• Graph Signal Processing

Background:

• Prostate cancer bone metastases cause significant complications, necessitating accurate imaging.
• Classical radiomics faces challenges with inter-scanner variability, segmentation dependence, and heterogeneity.
• A novel, robust radiomics approach is needed for reliable bone lesion characterization.

Purpose of the Study:

• Introduce a translational graph-based radiomics approach for quantitative descriptors.
• Enhance robustness and clinical reliability in imaging-based bone lesion analysis.
• Improve characterization of prostate cancer bone metastases.

Main Methods:

• Derived graph representation from PET/CT bone lesions using point clouds and Delaunay triangulation.
• Applied graph signal processing to extract orientation, connectivity, and transform-based features.
• Evaluated robustness against batch effects and segmentation variability; tested classification performance using LDA and SVM.

Main Results:

• Graph-based features captured non-redundant information and showed superior robustness.
• Models using proposed features outperformed classical radiomics in classification tasks.
• Achieved higher balanced accuracy and AUC with LDA and SVM compared to classical and combined features.

Conclusions:

• A translational graph-based radiomics framework offers improved robustness and diagnostic performance.
• The approach provides a promising tool for reliable PET/CT-based characterization of bone lesions.
• Enhances quantitative descriptors beyond conventional radiomics for prostate cancer management.