Deep Neural Recovery For Compressed Imaging

One of the biggest problem of MRI is its long scan time. To speed up the acquisition is possible to reduce the acquired points and reconstruct the missing ones after the acquisition. This work is based on a novel approach called LOUPE that tackles at same time the problem of reconstruction and the problem of finding the best under-sampling pattern. We contribute by adding some improvements to LOUPE. We introduce a regularization term, called "Flashback" that weights the difference of the under-sampled input with respect to its reconstructed version and improves the whole training performances. We show how Flashback can be also used as self-assessment to predict the reconstruction error of every scan at inference time. We introduce a method called "Back to the Future" that improves performances by swopping the acquired reconstructed frequency values with the original acquired frequency values. We introduce a layer called Trainable Inverse Fourier Transform (TIFT) that enables the classic layer for Inverse Fourier Transform to be trainable. Finally, we introduce "Prancing Pony", a method that automatically rises the slope of the sigmoid that produces the mask values, to automatically find its best value.