Nils Dienesch

Thesis Topic: Self-supervised deep learning for MR image reconstruction

Supervisor: Florian Knoll

Description

MRI is a popular medical imaging technique, but is inherently slow [1]. For this reason, accelerated MRI is important, in which images acquired at sub-Nyquist rates are reconstructed. Therefore, special techniques are necessary, because a simple reconstruction will lead to residual artifacts or noise. Recent reconstruction techniques focus on using deep learning methods to remove or reduce these.

Most of the current methods are trained using fully sampled reference images. However, fully sampled data is not available for alle imaging scenarios. One example is the acquisition of moving organs like the heart, where the image has to be acquired in a short period of time [1].

This encourages the investigation of self-supervised deep learning methods, which allow training without fully sampled reference data. In my master thesis especially the methods proposed in Yaman et al. [1] and Darestani et al. [2] are studied. The former uses learning properties of an under parametrized neural network for denoising the MRI reconstructions. Second divides the available k-space to get training and reference data from one under sampled acquisition.

[1] Yaman, Burhaneddin, Seyed Amir Hossein Hosseini, and Mehmet Akcakaya. “Zero-shot physics-guided deep learning for subject-specific MRI reconstruction.” NeurIPS 2021 Workshop on Deep Learning and Inverse Problems. 2021.

[2] Darestani, Mohammad Zalbagi, and Reinhard Heckel. “Accelerated MRI with un-trained neural networks.” IEEE Transactions on Computational Imaging 7 (2021): 724-733.