Reference or calibration scans are needed to reconstruct undersampled MRI data. Differences between reference scans and the undersampled acquisitions can exist, as they are acquired at different times and potentially during the presence of subject motion. This is more concerning in EPI acquisitions where geometric distortions, due to the long echo-spacings, are prevalent, and the effects of motion more problematic. Previous work has demonstrated that the use of a FLASH/GRE reference scan can significantly improve signal stability of single shot and segmented EPI acquisitions [1]. More recently, continuous segmented EPI acquisition (FLEET), has been proposed and reported to have superior tSNR to FLASH [2]. In segmented EPI, the duration between image segments of the same slice is long, allowing for signal recovery. This results in higher SNR but is more susceptible to segmentation artifacts. In FLEET, each segment is acquired sequentially with the minimum possible delay, reducing SNR, but also segmentation artifacts. Here we investigate the impact of the reference scan choice on tSNR and image quality in highly undersampled and slice accelerated 2D EPI data. Human subjects were scanned in accordance with the IRB at the University of Minnesota. Imaging was performed on a Siemens 7T Magnetom scanner equipped with a 32 channel nova-medical coil. Whole brain multiband acquisition was performed with 1.6mm^3 isotropic resolution. Each acquisition had 30 repetitions and 6 different reference scans: ‘FLASH 12’,’FLASH 52’,’FLEET 5’,’FLEET 20’, ‘Segmented EPI’, ‘Single-shot EPI’, with the numbers indicating nominal flip angles. Three different acquisitions were performed {MB=5, iPAT=2}, {MB=3, iPAT=3}, {MB=2, iPAT=4} with TR/TE {1000/22ms},{1400/22ms},{2000/22ms}. Scans were also acquired in the presence of subject motion and with/without B0 shim offsets. Additionally, 6 high-resolution scans were obtained with 0.9mm isotropic resolution, {MB=3, iPAT=3}, {TR/TE=3640/20ms}: three using a ‘FLASH 12’ and three using a ‘Fleet 20’ reference. Even/odd echo correction was performed using a three-line navigator and a linear fit of the phase. Image reconstruction was performed off-line in Matlab and compared with the online ICE reconstructions using the CMRR MB C2P sequence [3]. The different reference scans were applied to the same image data and time-series. A mask covering the brain was generated with FSL. tSNR was calculated for each voxel over the whole brain and then averaged. In addition to comparing the same data reconstructed with the different references scans, the effect of different acquisitions of the same reference scan type (FLASH 12 /FLEET 20) was also investigated.The impact of different reference scans on tSNR is shown in figure 1 for R=2, R=3, R=4, using both an optimal shim, and an off-resonance effect induced through an additional 40uT/m gradient offset along the X-axis. The tSNR calculated for different scans of the same data demonstrates that the variation between scans is larger than the differences between the FLASH and the FLEET methods. For FLEET 5, the SNR is low such that the unaliasing is inferior (data not shown). The signal for a midbrain slice for different experiments using FLASH 12 and FLEET 20 are shown in figure 2. This region had larger differences than more inferior or superior slices. For iPAT4, with induced off-resonance, a reduction in signal intensity can be seen on the top row, and, the tSNR (in figure 3) for those same slices shows a hyper-intensity, indicating increased fluctuations or reduced signal intensity. The stability between sequential scans is compared in figure 4, where $$$\Delta S $$$ is the change in signal between repeated acquisitions of the same reference scan type. The impact of the signal $$$ S $$$ is not visually clear, but the relative difference between FLASH and FLEET is noticeable.The results shown in figure 1 demonstrate that scan to scan variability is larger than any difference in tSNR between FLASH and FLEET. Further, overall image quality is also comparable, while ghosting is larger with segmented or single-shot EPI, compared to FLASH or FLEET. As such, the choice of either FLASH or FLEET would be ideal compared to an EPI calibration scan for in-plane undersampling. FLEET acquisitions may, however, be preferred as it takes about one third the time to acquire compared to FLASH. If time permits and similarity between successive acquisitions is important then FLASH might be preferable. There are also other factors yet to be considered, including the impact of tighter imaging FOVs and the effectiveness of the 3 line navigator.