Florian Fürnrohr

Thesis Topic: Zero-Shot Self-Supervised Reconstruction for 3D MRI Angiography

Supervisors: Marc Vornehm, Jens Wetzl (Siemens Healthineers), Daniel Giese (Siemens Healthineers), Florian Knoll

Description

The overall goal of the thesis is to implement a zero-shot learning procedure based on the paper of Yaman et al. [3], use it to reconstruct 3D accelerated MRI angiography data, and evaluate its performance against state-of-the-art techniques.

Yaman et al. [3] propose a deep learning approach for MRI reconstruction in a self-supervised manner, which enables scan-specific reconstruction without a database of fully-sampled measurements. The available k-space measurements for each scan are split into three disjoint sets: enforcing data consistency, defining a training loss, and enabling an early stopping criterion. An implementation of this method, that enables high quality 3D reconstruction of accelerated and partially measured (partial fourier) data, will be performed. The approach is model-agnostic, such that an examination of different architectures, like Residual Networks (ResNet) [2] or Variational Networks (VN) [1], will also be part of the thesis. An investigation of potential pretraining using cross-domain generalization is of interest, as the paper [3] suggests transfer learning to increase the reconstruction quality. The used Advanced MRA dataset consists of undersampled k-space measurements of the thorax for 3D angiography, performed using two echoes for Dixon fat-water separation. The redundancy of this information potentially increases stability in the reconstructions. In a first step, each echo will be used separately. In a second step, both echo measurements will be integrated into the reconstruction model, exploiting the relation between the data.

As we do not have fully-sampled measurements, the evaluation of the results is not straightforward. In the absence of a ground truth, quantitative measures (e.g. RMSE or SSIM) are obtained using samples with relatively low acceleration factors, artificially undersample these, and compare the reconstructions of our model with GRAPPA-reconstructions of the data without further undersampling. If necessary, expert ratings will be considered.

References

[1] K. Hammernik, T. Klatzer, E. Kobler, M. P. Recht, D. K. Sodickson, T. Pock, and F. Knoll. “Learning a Variational Network for Reconstruction of Accelerated MRI Data”. In: Magnetic Resonance in Medicine 79.6 (2018), pp. 3055–3071.
[2] K. He, X. Zhang, S. Ren, and J. Sun. “Deep Residual Learning for Image Recognition”. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016, pp. 770–778.
[3] B. Yaman, S. A. H. Hosseini, and M. Akcakaya. “Zero-Shot Self-Supervised Learning for MRI Reconstruction”. In: International Conference on Learning Representations. 2022.