ASL-to-PET Translation with Semi-Supervised Learning

Authors

Sahar Yousefi, Hessam Sokooti, Wouter M. Teeuwisse, Dennis F. R. Heijtel, Aart J. Nederveen, Marius Staring, and Matthias J. P. van Osch

Overview

PET can measure physiological function, but it requires radioactive tracers and can be costly. ASL is a non-invasive, non-radioactive MRI technique for measuring cerebral blood flow. This work investigates whether a neural network can translate ASL scans into PET images, while addressing the limited availability of paired ASL and PET training data.

Approach and evaluation

The semi-supervised multitask model learns from both paired ASL–PET scans and additional unpaired ASL scans. It uses a residual-based attention-guided mechanism to improve contextual features. The study also evaluates adding T1-weighted MRI as an input and uses seven-fold cross-validation followed by a double-blind observer study.

SSIM

0.85 ± 0.08

2D ASL + T1-weighted input

Mean squared error

0.01 ± 0.01

Reported image metric

PSNR

21.8 ± 4.5 dB

2D ASL + T1-weighted input

Project links