A Comparative Study of Deep Learning Methods for Brain Magnetic Resonance Image Reconstruction

Deep Learning shows a high promise in the field of neuroimaging with the recent development of models for data acquisition, classification problems, segmentation, and image synthesis and reconstruction. Magnetic resonance has been in recent times a very effective tool in the studies of various brain pathologies such as tumors and neurodegenerative diseases, however in the field of neurosciences reducing the patient’s exposure time has been very useful in patients who suffer from alterations in their nervous state whose movement can compromise the image quality for the execution of longer brain scan protocols. On the other hand, the high cost of high-field scanners has led to the development of portable low-field equipment with lower cost but lower performance that produce noisy images. In this work we present a comparative study between different techniques based on deep learning for image reconstruction in high and low field brain magnetic resonance images. We analyze methods based on convolutional networks, adversarial generative networks and propose a deep learning model for magnetic resonance image reconstruction based on the concepts of semantic genesis. The experiments developed in neuro-images taken by high and low field magnetic resonance scanners demonstrated a superior performance of the proposed architecture based on semantic genesis in terms of correlation and signal to noise ratio.