224 — MRI2MRI: A deep convolutional network that accurately transforms between brain MRI contrasts

Xiao et al (10.1101/289926)

Read on 01 April 2018
#neuroscience  #MRI  #CNN  #neural-network  #contrast  #imagery  #coregistration  #registration  #style-transfer 

MRI2MRI is a deep convolutional neural network that converts between contrast types of MRI scan.

Different types of MRI scan are used for different purposes, because they record different components of the underlying biology of the brain: The different contrast modes mean different things and highlight different parts of the anatomy.

This neural network generates images that correspond to alternative contrast modes: If you feed it a T1w MRI, it can output a T2-weighted version. This net leverages a concept similar to the style-transfer nets that were receiving so much attention recently for converting your webcam photos into Van Goghs.

Even if these images aren’t clinically relevant — and I’m sure no one (authors included) could be convinced they are — the authors explain that this system is still very useful for performing common actions like image coregistration: It is common to overlay several modalities of MRI in order to better understand the anatomy of the images with lower spatial resolution. This tool provides an image that much more closely matches the target of the registration, even if it does not look exactly the same. This means that the transformation matrix of the false MRI2MRI image can be applied to the true image to provide a better alignment.