Unsupervised deep learning framework for data-driven gating in positron emission tomography.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Nuclear Medicine

Background:

• Physiological motion, particularly respiratory motion, limits spatial resolution in positron emission tomography (PET) imaging.
• Motion-induced misregistration between PET and CT images can lead to attenuation correction artifacts.
• Respiratory gating techniques are crucial for mitigating motion artifacts and improving image quality.

Purpose of the Study:

• To propose a robust, data-driven approach for respiratory gating in PET imaging using an unsupervised deep clustering network.
• To leverage an autoencoder (AE) for extracting latent features from PET data to enable effective respiratory gating.

Main Methods:

• List-mode PET data were divided into short-time frames and reconstructed without corrections to minimize artifacts and reconstruction time.
• A deep autoencoder (AE) was trained on reconstructed frames to extract latent features for unsupervised respiratory gating.
• K-means clustering was applied to these latent features for respiratory gating, with evaluation against external signal and principal component analysis (PCA) methods.

Main Results:

• The proposed Deep Clustering method yielded gated PET images with superior contrast and sharper myocardium boundaries compared to external and Image PCA methods.
• Quantitative analysis demonstrated greater center of mass (COM) displacement and higher lesion contrast with the Deep Clustering method.
• The method effectively reduced motion-induced artifacts, improving diagnostic accuracy.

Conclusions:

• The proposed deep clustering framework provides superior respiratory gating performance for PET imaging.
• Validation with phantom and patient data confirmed the method's effectiveness over conventional techniques.
• This approach offers a promising solution for enhancing PET image quality in the presence of physiological motion.