Optimization-Based Image Reconstruction From Fast-Scanned, Noisy Projections in EPR Imaging

Tumor oxygen concentration image is essential to oxygen-image guided, precise radiation therapy. Electron paramagnetic resonance imaging is an advanced oxygen imaging technique. However, the scanning time is still comparatively long, leading to motion artifacts for static imaging and low time resolution for dynamic imaging. Usually, a projection signal at a specific angle is obtained by averaging thousands of repeatedly collected projections to suppress random noise and achieve a high signal-to-noise ratio (SNR). Reducing the repetition times of projection collecting at a specific angle may effectively speed up the whole scanning process. However, the EPR images reconstructed by the conventional three-dimensional filtered backprojection (FBP) algorithm from these fast-scanned, low SNR projections are too noisy to be used for further image postprocessing. In the paper, we investigate the capability of an optimization-based algorithm in accurate reconstruction from noisy projections. We designed a total variation constrained, data divergence minimization model, derived its Chambolle-Pock (CP) solving algorithm, and then validated and evaluated the CP algorithm via mathematical and physical phantoms. The studies show that the CP algorithm may accurately reconstruct EPR images from fast-scanned, noisy projections, and thus the whole scanning process may be speeded up four times compared with the full scan time demanded by the FBP algorithm in the image reconstruction of the complex physical phantom.