Deep Learning Techniques for Positron Emission Tomography Imaging

Deep learning has emerged as a transformative approach to enhance positron emission tomography (PET), addressing long-standing challenges of noise, limited resolution and lengthy acquisition times. Convolutional neural networks (CNNs) and generative adversarial networks (GANs) underpin a variety of end-to-end frameworks that can denoise low-dose scans, reconstruct full-dose quality images from reduced counts and accelerate dynamic PET protocols without sacrificing quantification. Cycle-consistent GANs and residual networks have been employed to synthesise standard-dose images from acquisitions that use a fraction of the usual radiotracer activity, while adaptive noise-aware architectures estimate per-scan noise levels to guide restoration. Deep image prior methods leverage the inherent bias of network structure in the absence of large training databases, enabling unsupervised denoising of dynamic sequences. Joint reconstruction techniques fuse PET with complementary anatomical data—such as magnetic resonance—via shared priors or kernel methods, sharpening functional boundaries and reducing artefacts. Recent multicentre studies demonstrate the generalisability of these algorithms across different scanner models, tracers and patient populations, preserving critical metrics such as standard uptake value (SUV) while reducing radiation dose or scan duration. Collectively, these advances promise to widen access to PET by lowering costs, enhancing patient comfort and unlocking new possibilities for real-time molecular imaging in both clinical and research settings.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Deep Learning Techniques for Positron Emission Tomography Imaging publication trend

The graph below shows the total number of articles in deep learning techniques for positron emission tomography imaging across all publications each year (not limited to Nature Index journals).

Technical terms

Convolutional Neural Network (CNN): A deep learning model employing convolutional filters to extract hierarchical spatial features from image data.

Generative Adversarial Network (GAN): A dual‐network framework in which a generator creates images and a discriminator evaluates their realism, enabling high-fidelity synthesis.

Cycle-Consistent GAN (CycleGAN): A GAN variant that enforces mapping consistency between two image domains, facilitating unpaired image translation.

Standard Uptake Value (SUV): A quantitative measure of radiotracer concentration in PET, normalised by injected dose and body metrics.

Deep Image Prior (DIP): An approach using the intrinsic structure of an untrained network as a regulariser for image reconstruction tasks without external training data.

Denoising: The process of reducing statistical noise in images while preserving critical anatomical or functional details.

Low-dose PET: PET imaging conducted with reduced radiotracer activity or shortened acquisition duration to minimise patient radiation exposure.

References

1. Adaptive 3D noise level‐guided restoration network for low‐dose positron emission tomography imaging. Interdisciplinary Medicine (2023).
2. Study of low-dose PET image recovery using supervised learning with CycleGAN. PLOS ONE (2020).
3. Deep learning-assisted ultra-fast/low-dose whole-body PET/CT imaging. European Journal of Nuclear Medicine and Molecular Imaging (2021).
4. Dynamic PET Image Denoising Using Deep Convolutional Neural Networks Without Prior Training Datasets. IEEE Access (2019).
5. Joint reconstruction of PET-MRI by exploiting structural similarity. Inverse Problems (2014).
6. MR-guided dynamic PET reconstruction with the kernel method and spectral temporal basis functions. Physics in Medicine and Biology (2016).
7. Low-count whole-body PET with deep learning in a multicenter and externally validated study. npj Digital Medicine (2021).