Anatomically aided PET image reconstruction using deep neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiological Physics

Background:

• Positron Emission Tomography (PET) imaging quality can be enhanced by integrating anatomical information from modalities like Computed Tomography (CT) and Magnetic Resonance Imaging (MRI).
• Iterative reconstruction frameworks offer potential for improving PET image quality but require effective regularization strategies.

Purpose of the Study:

• To propose a novel co-learning three-dimensional (3D) convolutional neural network (CNN) for improved PET image reconstruction.
• To integrate modality-specific features from PET/CT image pairs into an iterative reconstruction framework.
• To enhance PET image quality by leveraging anatomical information from CT scans.

Main Methods:

• A pretrained deep neural network was utilized to represent PET images, trained with low-count PET/CT pairs and high-count PET images as labels.
• Two CNN architectures (multichannel and multibranch) were investigated for integrating CT anatomical information into a constrained maximum likelihood (ML) framework.
• The proposed method was evaluated using Monte Carlo simulations and real patient datasets, compared against MLEM, kernel-based, and CNN-based methods.

Main Results:

• The proposed constrained ML reconstruction approach yielded superior image quality compared to existing methods.
• Lung tumors exhibited higher contrast with the proposed method than with CNN-based deep penalty reconstruction.
• Image quality was further enhanced by incorporating anatomical information, and liver standard deviation was reduced.

Conclusions:

• A supervised co-learning strategy effectively improves constrained maximum likelihood PET image reconstruction.
• The proposed method demonstrates an improved trade-off between lesion contrast and background standard deviation.
• This approach has the potential to enhance lesion detection in PET imaging.