Bayesian Uncertainty Estimation of Learned Variational MRI Reconstruction.

Recent deep learning approaches focus on improving quantitative scores of dedicated benchmarks, and therefore only reduce the observation-related (aleatoric) uncertainty. However, the model-immanent (epistemic) uncertainty is less frequently systematically analyzed. In this work, we introduce a Bayesian variational framework to quantify the epistemic uncertainty. To this end, we solve the linear inverse problem of undersampled MRI reconstruction in a variational setting. The associated energy functional is composed of a data fidelity term and the total deep variation (TDV) as a learned parametric regularizer. To estimate the epistemic uncertainty we draw the parameters of the TDV regularizer from a multivariate Gaussian distribution, whose mean and covariance matrix are learned in a stochastic optimal control problem. In several numerical experiments, we demonstrate that our approach yields competitive results for undersampled MRI reconstruction. Moreover, we can accurately quantify the pixelwise epistemic uncertainty, which can serve radiologists as an additional resource to visualize reconstruction reliability.

Using Synthetic Training Data for Deep Learning-Based GBM Segmentation.

In this work, fully automatic binary segmentation of GBMs (glioblastoma multiforme) in 2D magnetic resonance images is presented using a convolutional neural network trained exclusively on synthetic data. The precise segmentation of brain tumors is one of the most complex and challenging tasks in clinical practice and is usually done manually by radiologists or physicians. However, manual delineations are time-consuming, subjective and in general not reproducible. Hence, more advanced automated segmentation techniques are in great demand. After deep learning methods already successfully demonstrated their practical usefulness in other domains, they are now also attracting increasing interest in the field of medical image processing. Using fully convolutional neural networks for medical image segmentation provides considerable advantages, as it is a reliable, fast and objective technique. In the medical domain, however, only a very limited amount of data is available in the majority of cases, due to privacy issues among other things. Nevertheless, a sufficiently large training data set with ground truth annotations is required to successfully train a deep segmentation network. Therefore, a semi-automatic method for generating synthetic GBM data and the corresponding ground truth was utilized in this work. A U-Net-based segmentation network was then trained solely on this synthetically generated data set. Finally, the segmentation performance of the model was evaluated using real magnetic resonance images of GBMs.

A cryostatic, fast scanning, wideband NQR spectrometer for the VHF range.

In the search for a novel MRI contrast agent which relies on T1 shortening due to quadrupolar interaction between Bi nuclei and protons, a fast scanning wideband system for zero-field nuclear quadrupole resonance (NQR) spectroscopy is required. Established NQR probeheads with motor-driven tune/match stages are usually bulky and slow, which can be prohibitive if it comes to Bi compounds with low SNR (excessive averaging) and long quadrupolar T1 times. Moreover many experiments yield better results at low temperatures such as 77 K (liquid nitrogen, LN) thus requiring easy to use cryo-probeheads. In this paper we present electronically tuned wideband probeheads for bands in the frequency range 20-120 MHz which can be immersed in LN and which enable very fast explorative scans over the whole range. To this end we apply an interleaved subspectrum sampling strategy (ISS) which relies on the electronic tuning capability. The superiority of the new concept is demonstrated with an experimental scan of triphenylbismuth from 24 to 116 MHz, both at room temperature and in LN. Especially for the first transition which exhibits extremely long T1 times (64 ms) the and low signal the new approach allows an acceleration factor by more than 100 when compared to classical methods.