Electron Paramagnetic Resonance Imaging (EPRI) is a powerful tool for imaging oxygen distribution in tumor tissues¹. However, due to the extremely short T2 relaxation of electrons, single point imaging (SPI) is adopted in EPRI acquisition and leading to long scan time^1,2. In order to capture the dynamic changes of oxygen level inside tumors, EPRI acquisition with shorter temporal resolution is desired. We previously proposed use of Partial Fourier reconstruction to aid compressed sensing reconstruction for accelerated EPRI (PFCS)³. In this study, we explore the use of different sampling masks and phase map estimation methods to improve PFCS and achieve better image reconstruction.The PFCS method uses conventional CS sampling masks, where a small cube was fully sampled at k-space center (3% of k-space) with sparse sampling elsewhere (referenced as cube-sampling). A fully-sampled phase map (e.g. from the first temporal volume) was used to augment the original CS equation with fabricated measurements from conjugate symmetric signals³. Here we propose a new spherical-sampling pattern with a larger fully-sampled spherical mask in the center (12.5% of k-space) with the rest sparsely sampled (Fig. 1). In addition, to eliminate erroneous phase map estimation from a different time point, we use POCS⁴ (Projection Onto Convex Sets) method to estimate the phase map from the under-sampled data itself at each time point.All data were acquired using a home built EPRI scanner operating at 300MHz at a Zeeman field of 10.6mT. For phantom data, three tube phantoms containing 0%, 2%, 5% oxygen were imaged with 21x21x21 encoding steps. For in-vivo imaging, a 25g mouse with SCC7 tumor cells implanted into the left hind leg was imaged with 19x19x19 encoding steps. Alternating breathing cycles with air or carbogen (95% O₂ plus 5% CO₂) was applied to the mouse during the experiment to induce oxygen change in the tumor. Nine consecutive in vivo images were imaged during the breathing cycle with 4min temporal resolution for full k-space. Multi-gradients imaging process was applied to calculate oxygen map⁵. Data were then down-sampled with acceleration ratio of 4 (25% of k-space). Performances of PFCS with different sampling masks and phase maps were compared to conventional CS based on normalized mean square error (nMSE) with the true image reconstructed from the fully-sampled k-space.Figure 2 shows the comparison of the cube- vs. spherical-sampling masks for PFCS on tube phantom and in vivo mouse leg. Spherical-sampling preserves more image details and less blurring than the conventional cube-sampling. Figure 3 shows a comparison of different phase map estimation methods. Using POCS phase map, the images can be reconstructed even when using lower SNR data and results in a more accurate linewidth map compared to previous phase map estimator. Figure 4 gave the comparison of the reconstructed images using the new PFCS, PFCS, and CS techniques. The new PFCS retained most of the details of the true images than the other methods, and provided a correct estimation on linewidth map. Figure 5 shows temporal images from the left hind leg of the mouse subjected to alternating breathing of air and carbogen. Using the new reconstruction technique, the differences in the T2 with carbogen can be characterized with high temporal resolution. It is obvious the linewidth maps reconstructed by the POCS-based technique were able to capture the oxygen fluctuation induced by carbogen inhaling.We’ve demonstrated with the new PFCS, the nMSE of reconstructed images is reduced by more than 50%. The spherical-sampling mask outperforms the cube-sampling mask and retains the spatial resolution of fully sampled data. It is reasonable since the sphere-sampling mask contains a larger fully sampled center of k-space. This is especially important for images in low SNR such as in vivo EPRI. For phase map estimation, POCS gave a more accurate phase map estimation compared to the conventional method. This is possibly due to POCS estimates phase maps from the under-sampled data itself rather than using the phase map from the first time slot and prevent the estimation errors caused by phase difference between each time slot. The improvement in linewidths enables us to differentiate oxygenation and hypoxic regions and capture the oxygen change inside the tumor with only one-fourth of k-space and achieves a shorter temporal resolution (~ 1 min), which is an ideal temporal resolution for cycling hypoxia imaging⁶. Future work will continue to optimize the technique to obtain better temporal resolution to better capture information on cycling hypoxia.