Accelerated EPR Imaging Using Deep Learning Denoising

Motivation: Pulse EPR imaging (EPRI) is an excellent method to obtain pO2 maps in tissues, however, suffers from low signal-to-noise ratio (SNR) and image artifacts. Goal(s): In this project, our goal was to implement deep learning-based techniques to improve SNR and reduce artifacts. Approach: All EPRI experiments were performed using a 25 mT EPRI instrument, JIVA-25Ò. A UNet model, combined with Joint Bilateral Filters (JBF), was tested. Results: We demonstrate that UNet with 2-filter JBF provided the best outcome. Results showed that the model enabled a 10-fold faster acquisition. We demonstrate that the trained algorithm improves SNR in pO2 maps of mouse fibrosarcoma and kidneys. Impact: EPR images with physically enhanced deep learning techniques improve image SNR and reduce artifacts. This advancement can be translated to reduce acquisition time, reduce deposited power, and enable large object oxygen imaging, bringing EPRI one step closer to clinical translation.