Fast dynamic MRI plays an essential role for a variety of clinical applications such as dynamic contrast enhanced imaging, dynamic magnetic resonance angiography or dynamic imaging of motion or flow. Because MRI is an intrinsic slow technique, image acceleration is key for a high spatio-temporal resolution and short scan times. A variety of acceleration techniques has been suggested to address these limitations. Due to their linear reconstruction, techniques such as k-t GRAPPA¹, k-t SENSE² or its derivate k-t PCA³ are widely used for dynamic MRI applications. These techniques are conventionally based on a fully sampled central region of k-space and a constant undersampling of the higher frequencies on a sheared grid in the k-t domain, allowing for a fast and stable linear reconstruction process. However, variable density sampling in k-space, with high density in the k-space center to lower density in the k-space periphery, has been shown to be beneficial for fast imaging techniques such as compressed sensing⁴. Herein a new type of variable density k-t space sampling scheme for acquisitions with sheared grid k-t space sampling and an adapted reconstruction framework based on a linear reconstruction is proposed.In typical ‘linear’ k-t acquisitions, data is sampled on a sheared grid in the k-t domain with a fully sampled region in the central k-space, which provides the so called training data, and a constantly undersampled region in the k-space peripheral (Fig. 1A). The proposed method uses a modified sampling scheme, which includes a fully sampled central region of the k-t space and undersampling of the k-space periphery with variable density on multiple shells (Fig. 1B), resulting in constantly increasing undersampling factors (e.g. 2, 4, 8, 16, 32) while moving towards the k-space periphery. The sampling pattern can be divided into training data and the different undersampling factors covering different ranges of k-space (Fig. 2C). Data reconstruction is started on the most central undersampling shell which is reconstructed using k-t PCA resulting in a low resolution image. The next undersampling layer is then reconstructed using the reconstructed images from the previous shell as training data. This process is repeated until the most outer undersampling shell is reached and the final images are reconstructed. A scheme of the proposed reconstruction framework is shown in Fig. 2. To evaluate the proposed method, a numerical phantom for dynamic contrast enhanced imaging was designed using signal curves as based on a Tofts model with a maximal signal to noise ratio of 50. Data was undersampled on a sheared grid as used for convention k-t type of sequences (Fig. 1A) and using the proposed method with variable k-space density (Fig. 1B). Data was reconstructed according to the scheme shown in Fig. 2, based on a k-t PCA reconstruction.Fig. 3A shows a single time point of the dynamic numerical phantom (Fig. 3 left panel) and reconstructions using k-t PCA (Fig. 3 middle panel) and the proposed method (Fig. 3 right panel) from the same time point for net acceleration factors of R_eff = 7.7 and 8.0, respectively. Time intensity plots are shown for the different reconstructions along with time-intensity profiles and their error. Fig. 4 shows the same data as Fig. 3 for acceleration factors of R_eff = 15.8 and 16.1.The reconstructed dynamic curves and spatio-temporal plots for R_eff = 7.7 and 8.0 show good agreement to the reference for the k-t PCA and the proposed method with only slight differences. For higher acceleration factors of R_eff = 15.8 and 16.1 the advantage of the proposed multi shell variable k-space method becomes apparent. The dynamic curves reconstructed with the proposed method show good agreement to the reference curves, while the k-t PCA reconstruction shows strong deviations (temporal blurring) from the reference curves (Fig 4B). This advantage of the proposed method over the conventional k-t PCA reconstruction is also apparent it the spatio-temporal signal variations shown in Fig. 4C, despite the slightly higher acceleration factor. This emphasizes the benefit of the proposed multi shell variable k-space density method for highly accelerated dynamic MRI.The herein prosed method for fast dynamic MRI is based on a multi shell variable k-space sampling in combination with a linear reconstruction. The proposed method is shown to be comparable to conventional k-t PCA for moderate acceleration factors while outperforming conventional k-t PCA for high accelerations factors allowing for ultra-fast dynamic MRI. Future work will concentrate on prospectively accelerated data acquisition using the proposed method.