Can PET FDG Deep Texture Representations Be Made Robust?

The efficacy of Convolutional Neural Networks (CNNs) in recognizing cancerous textures in PET scans is well acknowledged, yet model generalizability across different cancer types, imaging acquisitions, and reconstruction protocols remains largely unexplored. This study evaluates a 3D CNN originally trained on esophageal cancer data, applied to lung cancer and lymphoma. We test the model using organ-specific PET data from four glycolytic volumes with high Standardized Uptake Values (SUV), acquired from different scanners and acquisition protocols. Through Principal Component Analysis (PCA) of the network’s sub-logit layer activations, we demonstrate its ability to distinctly identify textural features across different cancers, scanners and acquisition protocols, highlighting its robust cross-cancer applicability and potential for enhancing diagnostic precision.