Figure 1
Proposed pipeline. Experiments in proposed study entail generating ground truth T2 maps from MAPSS, simulating accelerated acquisition of T2-weighted MAPSS images, and training a network to predict T2 maps from undersampled images. (1) MAPSS contains 7 images, 3 that are T2 weighted, 3 T1ρ weighted, and 1 shared; the T2 and shared image weightings are extracted, registered, and fitted slice-wise to yield ground truth T2 maps. To simulate accelerated acquisition, each volume of coil-combined magnitude T2 weighted images acquired at a given echo time are Fourier transformed, undersampled along the ky–kz plane with a center-weighted Poisson disc pattern, and inverse Fourier transformed to yield a simulated accelerated acquisition of a volume. Finally, undersampled T2 weighted images acquired at all echo times for the same anatomic slice are concatenated and fed to the proposed recurrent UNet architecture, which predicts the T2 map appearance for the slice. Training is done slice-wise with a multi-component loss function that includes a novel ROI-specific L1 loss that optimizes predicted T2 maps in cartilage and IVD ROIs, with other components that improve training stability and encourage retention of textures.
