Group equivariant pyramid network for respiratory motion correction on PET image

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Respiratory motion causes significant blurring in PET images, hindering accurate lesion detection and diagnosis.
• Current artefact correction techniques lack advanced feature learning, limiting their effectiveness in removing motion-induced blurring.

Purpose of the Study:

• To develop a novel deep learning framework for robust respiratory motion artefact correction in 3D PET imaging.
• To enhance the clarity and diagnostic reliability of PET images affected by patient movement during scans.

Main Methods:

• Development of a Group Equivariant Dual-Pyramid Network (GEPN) utilizing SE(3)-equivariant convolutions for 3D PET data.
• Integration of attention-enhanced CNNs and SE(3) G-CNNs for efficient organ feature extraction.
• Implementation of a Lie group-based motion decomposition strategy in the decoder for inter-organ displacement and SE(3) component resolution.

Main Results:

• GEPN demonstrated superior performance across geometric, lung phantom, and clinical PET datasets, achieving high Dice coefficients (e.g., 81.01% on clinical data).
• The framework significantly improved artefact correction compared to baseline models, enhancing lesion clarity and motion pattern alignment.
• Quantitative and perceptual metrics confirmed GEPN's effectiveness in mitigating respiratory motion artefacts.

Conclusions:

• The GEPN framework successfully addresses respiratory artefacts in PET imaging through a unique combination of group-equivariant architectures and attention mechanisms.
• GEPN enhances PET image quality, offering a robust solution for improved diagnostic reliability and clinical decision-making.
• The method provides sharper PET images, aiding in precise tumor delineation for improved radiotherapy planning and treatment assessment.