Ultra-Fast / Low-Dose PET Imaging Using Transformers

Overview

This collaborative medical-imaging project explored whether Transformer-based deep learning could recover full-dose-like PET image quality from substantially reduced PET acquisition data. The goal was to support lower administered radiotracer dose or shorter acquisition time while preserving clinically useful image quality.

The study used clinical ¹⁸F-Flutemetamol brain PET/CT list-mode data. Low-dose PET images were retrospectively simulated using only 5% of the acquired list-mode events, then translated toward full-dose image quality using a hybrid Vision Transformer architecture combining Transformer attention blocks and convolutional blocks.

My role

Co-author and contributor to the deep-learning model design and manuscript development. My main contribution focused on designing the low-dose-to-full-dose PET image-translation model, particularly the Transformer-based architecture combining attention and convolutional blocks. I also contributed to writing the first draft of the paper, with emphasis on the deep-learning methodology and model-design sections.

Methods & tools

The project used 76 clinical PET/CT images acquired on a Biograph mCT PET/CT scanner in list-mode format. Low-dose counterparts were simulated by reconstructing 5% of the PET list-mode events. A hybrid encoder–decoder architecture combining Vision Transformer attention blocks with convolutional blocks was implemented to predict full-dose PET images from low-dose inputs in image space.

Model performance was evaluated using standard quantitative image-quality metrics, including structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), root mean squared error (RMSE), correlation, and joint-histogram analysis against reference full-dose PET images.

Key findings

The Transformer-based model substantially improved image quality compared with low-dose PET images. In the reported evaluation, predicted full-dose images achieved higher SSIM, PSNR, and correlation, while RMSE decreased compared with low-dose images. Visual assessment also showed effective noise suppression while preserving underlying ¹⁸F-Flutemetamol uptake patterns.

Why it matters

Reducing injected activity or acquisition time in PET imaging can improve patient comfort and reduce radiation exposure, but usually leads to noisier images and reduced diagnostic value. This project demonstrated a proof-of-concept deep-learning approach for recovering full-dose-like PET image quality from highly reduced acquisition data.

Related output

Conference paper / poster

Ultra-Fast / Low-Dose PET/CT Imaging Using Transformers

IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), 2022 · Co-author · IEEE Xplore

See all Publications.