Compressed Sensing Inspired Neural Decoder for Undersampled MRI with Self-Assessment

An important problem in magnetic resonance imaging (MRI) is the long time lapse required to acquire a fully sampled, high resolution scan. To speed up acquisition, Compressed Sensing (CS) has been used and recently coupled with Neural Networks (NN). In the latter setting, commonly CS has been split into two different problems: i) design of the encoder, or selection of the undersampling pattern, and ii) design of the decoder. A significant progress was recently introduced by a solution (called LOUPE) where encoding and decoding are simultaneously addressed. Here we propose an improvement of this model, called 'regularized-LOUPE' (r-LOUPE), which add measurement constraint into the picture, resulting in a ×8 speed-up in the MRI acquisition time. A further benefit of our methodology is that measurement constraint can be leveraged to implement a self-assessment tool able to predict the reconstruction error and to identify possible out-layers.