Recovery of the spatially-variant deformations in dual-panel PET reconstructions using deep-learning

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Image Reconstruction

Background:

• Dual-panel PET scanners, like Breast-PET, produce spatially variant image deformations due to limited-angle data and depth-of-interaction effects.
• Previous work utilized time-of-flight (TOF) and point-spread function (PSF) models to mitigate these deformations, improving small lesion quantification but limited for larger objects.

Purpose of the Study:

• To investigate the efficacy of deep learning (DL) networks in correcting severe, spatially-variant image deformations in dual-panel PET systems.
• To assess DL's ability to improve quantitative performance in the presence of limited-angle reconstruction artifacts.

Main Methods:

• Simulated PSF deformations in a generic dual-panel PET system's image space.
• Trained DL networks on simulated data with ground truth, then tested on simulated and acquired dual-panel Breast-PET phantom data.
• Utilized DIRECT-RAMLA reconstructions as input for the DL network on both synthetic and real B-PET data.

Main Results:

• DL networks demonstrated a significant capacity to eliminate image deformations inherent in limited-angle PET systems.
• The deep learning approach substantially improved the quantitative accuracy of reconstructed images.

Conclusions:

• Deep learning offers a powerful solution for correcting complex image deformations in dual-panel PET imaging.
• DL-based methods can significantly enhance the quantitative performance of Breast-PET scanners, overcoming limitations of traditional PSF modeling.