Structural imaging at high field can be performed at higher resolution but comes with increased scanning time compared to lower field. 3D-T1 weighted images of the human show a high autocorrelation due the limited variation in T1 within gray and white matter. This sparsity can be exploited to reduce scanning time by sparse sampling and compressed sensing (CS) reconstruction¹. Here, we aim to increase spatial resolution while maintaining scanning time for 3D-T1 weighted imaging by on-scanner undersampling at 7T.A consenting healthy volunteer was scanned on a 7.0T scanner (Achieva, Philips Healthcare, Cleveland, USA) equipped with a 32-channel Nova head coil.Four 3D T1-weighted Turbo Field Echo (TFE) sagittal imaging series were acquired, with FOV 246x246x174 mm, TFE factor 352, radial turbo direction, elliptical k-space shutter, inversion time 1300 ms, shot interval 4500 ms, repetition time 5.2 ms, echo time 2.2 ms, flip angle 7 degrees, bandwidth 502 Hz and auto shimming.Scan 1 was made at 0.5mm³ resolution, and 4 times CS undersampled (within the elliptical kspace shutter) using a variable density poisson disc and an elliptical fully sampled center with a short axis length of 60 (see Fig. 1). TFE shots were sampled in a radial fashion in the k(y,z)-plane (Fig. 1, right), without slice oversampling (SoS). Scan 2 was made at 0.7mm³ resolution, with SENSE undersampling 1.6x1.5 in the AP- and RL-direction, and SoS factor 1.4. Scans 3 and 4 were made at 0.5 mm³ resolution, with SENSE undersampling 1.9x1.8 and 2.2x2.2 respectively, and SoS factors 1 and 1.4.Scanning time was 9m10s for scans 1, 3 and 4 and 8m30s for scan 2. Scans 2, 3 and 4 required an SENSE reference scan which was acquired with a scan duration of 1m4s, so that scan 1 had the shortest total scanning time.SENSE data was reconstructed online. Scan 2 was linearly upsampled from 0.7mm³ to 0.5 mm³ to enable visual comparison with scan 1. CS data of scan 1 was reconstructed offline using the Berkeley Advanced Reconstruction toolbox². After a FFT along the frequency axis, reconstruction was performed in 2D along the undersampled phase encoding axes. Coil sensitivity maps were estimated from the data³. Reconstructions were L1-regularized with a wavelet sparsity transform. A regularization parameter was heuristically set to 0.05 to balance between spatial blurring and noise levels.An observed slight misalignment between scans due to head motion was not corrected for to avoid interpolation effects in the data.The CS-reconstructed scan 1 shows no streaking artifacts as can be observed in SENSE scans 2 and 3 due to bright signal in the vasculature (Fig. 2). Also, high resolution SENSE data (scans 3,4, Fig. 2) is medially more affected by noise. Scan 1 shows better contrast between gray matter and CSF (Fig. 3). Furthermore, partial volume effects in small gyri are diminished (Fig. 4). With the chosen regularization value, the visually perceived noise level is acceptable when navigating through a stack of slices (Fig. 5).The chosen sparse sampling scheme provides sufficient SNR by more densely sampling of the center of k-space. The sampling density at higher frequencies is sufficiently high to minimize PSF-blurring while keeping scanning time within 10 minutes. Results show that contrast is indeed maintained as was aimed for with the sampling strategy in Fig. 1b. The absence of streaking artifacts in Fig. 2 might be explained by the adopted ESPIRiT autocalibration method³, resulting in a spatial dispersion of artifacts which appear more local in SENSE-reconstucted data. Careful observation of the CS-reconstruction in Fig. 2 shows stronger regularization (less noise) in rows of data contain bright vessels. The varying noise level throughout the slice needs to be accounted for by the human reader or processing algorithms. The wavelet sparsity transform nulls signal values close to the noise floor, effectively removing any CSF-signal (Fig. 3). Signal intensity profiles in Fig. 4 show partial voluming in scan 2 and insufficient SNR in scans 3 and 4 for accurate size estimation of small gyri. The resulting noise originating from aliasing and scanner noise visually appears more sharp, but is regularized to a level to obtain sufficient image quality.A limitation of this work is the long reconstruction time, caused by data traffic (10Gb of raw data) and the current CPU reconstruction which runs overnight.Adaptive sparse sampling enables increasing spatial resolution while maintaining scanning time in 3DT1-weighted TFE imaging of the human brain at 7 Tesla.