Spatial Normalization of MRI Brain Studies Using a U-Net Based Neural Network

Abstract

Over recent years, Deep Learning has proven to be an excellent technology to solve problems that would otherwise be too complex. Furthermore, it has seen great success in the area of medical imaging, especially when applied to the segmentation of brain tissues. As such, this work explores a possible new approach, using Deep Learning to perform spatial normalization on Magnetic Resonance Imaging brain studies. Spatial normalization of Magnetic Resonance images by tools like FSL, or SPM can be inefficient for researches as they require too many resources to achieve good results. These resources include, for example, wasted human and computer time when executing the commands to normalize and waiting for the process to finish. This can take up to several hours just for one study. Therefore, to enable a faster and easier method to normalize the data, a U-Net based Deep Neural Network was developed using Keras and TensorFlow. This approach should free the researchers’ time for other more relevant tasks and help reach conclusions faster in their studies when trying to find patterns between the analyzed brains. The results obtained have shown potential by predicting the correct brain shape in less than 10 s per exam instead of hours even though the model did not yet accomplish a fully usable spatial normalized brain.