Anatomically and metabolically informed diffusion for unified denoising and segmentation in low-count PET imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Positron emission tomography (PET) image denoising and segmentation are crucial for diagnosis but often treated independently.
• Existing methods overlook the synergistic benefits between denoising and segmentation in PET analysis.
• Low-count PET imaging presents challenges due to increased noise and reduced image quality.

Purpose of the Study:

• To introduce the anatomically and metabolically informed diffusion (AMDiff) model, a unified framework for PET denoising and lesion/organ segmentation.
• To leverage the inherent synergies between denoising and segmentation tasks for improved performance in low-count PET imaging.
• To enable direct quantification of clinical metrics, such as total lesion glycolysis (TLG), from low-count PET data.

Main Methods:

• Developed AMDiff, a unified framework integrating a semantic-informed denoiser (diffusion strategy) and a denoising-informed segmenter (nnMamba architecture).
• Incorporated a lesion-organ-specific regularizer to constrain denoised outputs and a denoising revision module to enhance segmentation.
• Utilized a warming-up mechanism to optimize multi-task interactions within the AMDiff model.

Main Results:

• AMDiff demonstrated superior performance on multi-vendor, multi-center, and multi-noise-level datasets.
• For low-count PET data (below 20% clinical levels), AMDiff achieved TLG quantification biases of -21.60±47.26%.
• AMDiff outperformed standalone denoising and segmentation methods, reducing normalized root-mean-square error by 22.92% (lesion) and improving Dice coefficients by 10.17% (lesion).

Conclusions:

• The AMDiff model effectively integrates PET denoising and segmentation, surpassing the performance of independent methods.
• This unified framework significantly improves the analysis of low-count PET imaging, enabling accurate quantification of key clinical metrics.
• AMDiff offers a promising advancement for PET-aided diagnosis by exploiting multi-task synergies.