Unsupervised quantitative MRI reconstruction with densely-connected generative modeling and Bloch subspace projector

Motivation: The physical model of quantitative MRI varies with different applications and subjects, which brings difficulties for reconstruction algorithms. Goal(s): To propose an unsupervised reconstruction algorithm for accelerated quantitative MRI with arbitrary physical model. Approach: We propose a densely-connected generative architecture to model the quantitative MRI images. Besides, we adopt the Bloch subspace modeling to incorporate the physical model into the algorithm. We devise an ADMM algorithm for MR-physics-constrained optimization of the reconstruction. Results: Preliminary study was conducted on brain T2-mapping data. The proposed method achieves the state-of-the-art performance compared with TV, low-rank, supervised and unsupervised methods, without the need for any fully-sampled data. Impact: The proposed reconstruction framework is flexible with different physical models. It has potential to be used for other quantitative MRI imaging applications.