As OXO71-based Electron Paramagnetic Resonance Oxygen Imaging (EPROI) transitions from small to large animal models, and ultimately to humans, various challenges must be addressed. Long acquisition times and low SNR of individual scan pose significant challenges.
In the paper "Accelerated EPR Imaging using Deep Learning Denoising" , published in Magnetic Resonance in Medicine journal, in collaboration with Dr. McMillan from University of Wisconsin-Madison, we tried to tackle some of these issues using deep learning techniques. Our approach leverages a UNet enhanced with residual units, followed by two trainable joint bilateral filters (UNet+JBF2). The results are extremely encouraging:
- 10× reduction in number of averages (from 150 to just 15 shots) while maintaining SNR
- Noise levels comparable between the filtered 15-shot acquisition and the original 150-shot acquisition (see Figures d, e, below).
- 3x improvement in image quality for 150-shots acquisitions, boosting the SNR from 240 to 666 (p-value = 0.0074).
Even more excitingly, our model generalizes to oxygen maps. It successfully denoises amplitude maps without distorting pO2 estimations (Figure a, below) and enhances SNR by 2.4 fold for in vivo tumor pO2 maps (Figure b, below).
Click here or on the images to learn more!
