MDPET

MDPET performs unified motion correction and denoising of low-dose gated positron emission tomography (PET) images using an adversarial network to produce motion-compensated reconstructions.

Key Features:

• Unified motion correction and denoising: Simultaneous motion estimation and noise reduction within a single framework for low-dose gated PET.
• Adversarial network: Adversarial training is used to refine reconstruction quality and denoising performance.
• Temporal Siamese Pyramid Network (TSP-Net): TSP-Net integrates spatial and temporal modeling specifically for gated PET sequences.
• Siamese Pyramid Network (SP-Net): SP-Net captures spatial hierarchies and extracts robust features from PET data to aid motion estimation.
• Recurrent motion estimation layer: A recurrent layer estimates motion across different gates and tracks temporal motion patterns.
• Motion-compensated PET reconstructions: Produces high-quality, motion-compensated PET images from low-dose gated data.
• Validation on human datasets: Experimental validation and comparative analyses on human PET datasets demonstrated improved motion estimation and denoising.

Scientific Applications:

• Low-dose gated PET reconstruction: Improve image quality and motion compensation in low-dose gated PET reconstructions.
• Motion estimation from low-count gated images: Estimate respiratory motion directly from low-dose gated PET frames.
• Denoising of low-dose PET: Reduce noise in low-injection-dose gated PET sequences while preserving temporal motion information.
• Clinical research and diagnostic PET imaging: Support research and diagnostic workflows that require motion correction under reduced radiation dose.

Methodology:

Unified adversarial network employing a Temporal Siamese Pyramid Network (TSP-Net) composed of a Siamese Pyramid Network (SP-Net) and a recurrent layer to estimate motion across gates, with adversarial training jointly optimizing denoising and motion correction to produce motion-compensated PET reconstructions.