Susceptibility artifacts and image distortions are a major problem in diffusion-weighted EPI for various clinical applications. The problem may be overcome by the use of single-shot STEAM (ssSTEAM) MRI [1] which allows for anatomically correct imaging without sensitivity to magnetic field inhomogeneities due to the acquisition of radiofrequency-refocused echoes. Because multiple 180˚-pulses as in spin-echo MRI are not required, exceedingly high energy absorption rates as well as artifacts due to violations of the Carr-Purcell-Meiboom-Gill condition are also avoided. Additionally, the black-blood property of the sequence [2] suppresses unwanted signal contributions to the diffusion-weighted signal. So far, a poor signal-to-noise ratio (SNR) has been the main limitation for a more extended clinical use of diffusion-weighted ssSTEAM MRI, although high undersampling factors with correspondingly increased readout flip angles should improve the SNR. Here, we demonstrate the feasibility of undersampled radial diffusion-weighted ssSTEAM MRI of the human brain in combination with image reconstruction by the iteratively regularized Gauss-Newton method [3].The diffusion-weighted ssSTEAM sequence (Fig. 1) prepares the magnetization with a diffusion-encoded spin echo, followed by a 90˚ pulse. A series of small angle pulses generates stimulated echoes, one for each radial spoke of the sampling trajectory. A fat saturation pulse is applied before each slice. Multiple slices without gaps are acquired in an interleaved manner using a sufficiently long TR. The procedure is repeated for each b-value and each diffusion direction. All studies were performed at 3 T (Magnetom Prisma, Siemens Healthcare) using the standard 64-channel head coil. Validation measurements were performed for the diffusion coefficient of distilled water using b-values of 0, 100, …, 1000. The imaging protocol for healthy volunteers contained the following parameters: FOV: 192 × 192 mm², in-plane resolution: 1.5 mm, slice thickness: 5 mm, sampling: 25 spokes, TE: 7.44 ms, spin-echo-TE: 36.1 ms, TRS: 6.89 ms, TR: 12 s, TM: 10 ms. With these parameters, one image with b = 0 and 6 diffusion-weighted images in different directions with a b-value up to 1000 can be acquired in 84 s for 40 slices covering the entire brain. Image reconstruction was based on the iteratively regularized Gauss-Newton method [3] modified to jointly estimate the image content and coil sensitivities of the entire volume, i.e. all slices. The original image regularization, penalizing the L2-norm of the estimated data, is replaced by a regularization which penalizes high values of $$$h(ρ,c)=‖ρ‖_2^2+∑_{i=2}^{s-1}‖c_i-0.5β(c_{i+1}+c_{i-1})‖_2^2 $$$ with ρ the image content and c_i the coil sensitivity profile in slice i, β is a regularization parameter. The procedure takes advantage of the spatial smoothness of coil sensitivities perpendicular to the image section. Post-processing involved a novel image denoising filter using a modified non-local means algorithm [4].The diffusion coefficients measured in distilled water were in agreement with literature values [5], see Table 1. In vivo experiments yielded anatomically correct diffusion-weighted images of the brain with acceptable SNR and no averaging (Fig. 2). No further post processing (e.g., for removal of artifacts) was necessary. The present results demonstrate that undersampled radial ssSTEAM MRI, in combination with IRGNM-based image reconstruction, is a potentially useful strategy for clinical diffusion-weighted MRI, e.g. for tumor and stroke diagnosis. Its performance will be further tested in studies of the brain as well as other organs such as the liver to explore the range of suitable clinical applications.