Evaluation of deep learning-based scatter correction on a long-axial field-of-view PET scanner

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence in Healthcare

Background:

• Long-axial field-of-view (LAFOV) PET systems offer higher sensitivity but increase scatter, complicating image quality and quantification.
• State-of-the-art scatter correction methods like single scatter simulation (SSS) face limitations with LAFOV systems.
• Accurate scatter correction is crucial for reliable image reconstruction and quantitative analysis in advanced PET imaging.

Purpose of the Study:

• To evaluate the performance of deep learning-based scatter estimation (DLSE) for scatter correction in LAFOV total-body PET.
• To compare the DLSE method against single scatter simulation (SSS) for scatter estimation accuracy and robustness.
• To assess the impact of DLSE on lesion contrast recovery and overall image quality in clinical PET datasets.

Main Methods:

• A convolutional neural network (CNN) U-Net architecture was employed for DLSE, utilizing emission and attenuation sinograms.
• The DLSE network was trained using Monte Carlo simulations of XCAT phantoms with [F]-FDG PET acquisitions on a Siemens Biograph Vision Quadra scanner.
• Method performance was validated on simulated data and seven clinical [F]-FDG and [F]-PSMA PET datasets, comparing DLSE to MC ground truth and SSS.

Main Results:

• DLSE demonstrated superior accuracy and robustness to variations in patient size and dose compared to SSS on phantom data.
• The DLSE method achieved better lesion contrast recovery in phantom studies.
• Clinical evaluation showed improved lesion contrasts for [F]-FDG datasets and consistent performance for [F]-PSMA datasets, even without specific PSMA training.

Conclusions:

• The proposed DLSE method accurately estimates scatter in LAFOV PET systems using raw data.
• DLSE offers a more robust and accurate alternative to SSS for scatter correction in total-body PET.
• This deep learning approach holds significant potential for enhancing quantitative accuracy and image quality in advanced PET imaging.