Bob van Hoek^{1}, Kai Lønning^{2,3}, Hanneke Hulst^{4}, Frans Vos^{1,5}, and Matthan Caan^{6}

Deep learning can accelerate MRI beyond what is currently possible. Broad clinical application requires generalizability to multiple contrasts, acceleration levels and pathologies. Here we explore how a Recurrent Inference Machine trained on healthy volunteer T1-weighted brain images performs in such a situation, by reconstructing FLAIR images with white matter lesions, in simulation and prospectively undersampled patient data. Lesion contrast is maintained up to 6x acceleration and higher than in compressed sensing (CS) reconstruction, and all lesions are retained compared to CS.

- Hammernik K, Klatzer T, Kobler E, Recht MP, Sodickson DK, Pock T, Knoll F (2018): Learning a variational network for reconstruction of accelerated MRI data. Magn Reson Med 79:3055–3071.
- Putzky P, Welling M. Recurrent Inference Machines for Solving Inverse Problems. arXiv:1706.04008, 2017.
- Uecker M, Lai P, Murphy MJ, Virtue P, Elad M, Pauly JM, Vasanawala SS, Lustig M. ESPIRiT — An Eigenvalue Approach to Autocalibrating Parallel MRI : Where SENSE Meets GRAPPA. Magn. Reson. Med. 2014;1001:990–1001.

Figure 1. A Recurrent Inference Machine (RIM) was trained on 10 non-accelerated T1-weighted scans. Training successfully converged, illustrated by a reconstructed T1w-scan (lower-left). After that, FLAIR data with simulated lesions were reconstructed using the trained RIM. Lesions had varying intensity achieved by multiplying by values (1+a) with a=[0…100%], and are bandwidth-limited by convolving with a Gaussian kernel with $$$\sigma$$$=1px. Data were accelerated over a range of 4-10x.

Figure 2. The Recurrent Inference Machine (RIM) unrolled in time. The network parameters are shared over time steps. Based on the current estimate of the image $$$x_t$$$, the log-likelihood gradient $$$\nabla_{y|x_t}$$$ is computed based on the forward model with $$$c$$$ coil sensitivities $$$S$$$ and undersampling pattern $$$P$$$. The network $$$h_\phi$$$ contains two GRU’s embedded in three CNNs and two hidden states $$$s$$$. The image update $$$\Delta x_t$$$ is add to obtain the new estimated image.

Figure 3. Illustration of a simulated lesion at intensities of 30% and 60% in the original FLAIR slice, with reconstructions using the Recurrent Inference Machine (RIM) and Compressed Sensing (CS). The lesion appears as a small punctate object, adjacent to cortical GM.

Figure 4. Lesion contrast as a function of simulated lesion intensity for the original scan (1x orig), the Recurrent Inference Machine (RIM) and Compressed Sensing (CS) for two modelled lesions in deep and cortical WM.

Figure 5. Reconstructed images of 6x prospectively undersampled data of a patient with a mild white matter lesion load. Note that all lesions remain visible in the RIM-reconstruction compared to the CS-reconstruction.