October 2025 - New AI Model “AMDiff” Boosts PET Image Quality and Lesion Segmentation

UC Davis collaborators at Yale University have developed AMDiff, a groundbreaking artificial intelligence model that simultaneously cleans up noisy PET scans and identifies organs and lesions – tasks traditionally handled separately. By combining these functions into a unified framework, AMDiff takes advantage of the natural link between image clarity and anatomical accuracy. 

Yale postdoc Menghua Xia developed AMDiff using advanced diffusion-based denoising mechanism and novel nnMamba segmentation architecture. AMDiff produces clearer images and more precise measurements, even from scans with very low tracer counts. In multi-center tests, the AMDiff showed over 20% lower image error and notably higher lesion segmentation accuracy than the baseline single-task methods. 

The model also enables computation of key clinical metrics like total lesion glycolysis directly from low-dose scans, potentially reducing radiation exposure and improving diagnostic confidence across diverse imaging systems.