Resource-Aware Conditional Diffusion for CT-to-PET Translation Supporting Rural Oncology Imaging

Abstract Access to positron emission tomography (PET) remains limited in rural and low-resource healthcare settings due to high infrastructure cost and radiotracer logistics. This restricts early oncologic screening in underserved populations. The study proposes a rural-optimized conditional diffusion framework for synthetic PET generation directly from widely available CT scans. The architecture employs a two-stage residual design consisting of a lightweight coarse predictor followed by computationally efficient diffusion refinement with reduced timesteps and deterministic sampling. A multi-objective SUV-aware loss ensures metabolic consistency. To emulate rural deployment conditions, this study simulates low-dose noise, Hounsfield unit miscalibration, and resolution degradation. Clinical validation demonstrates strong structural fidelity (SSIM 0.81) and stable SUVmean preservation. Domain-matched training achieves SUVmax error as low as 0.61. Cross-dataset analysis highlights the importance of SUV harmonization for robust rural deployment. This work presents a resource-sensitive AI frame-work supporting equitable oncology screening in rural healthcare systems. Highlights Two-stage residual conditional diffusion for CT-to-PET translation. SUV-aware multi-objective optimization preserves metabolic biomarkers. Few-shot adaptation improves cross-dataset SUV calibration.