Home
Welcome to the MRH-Net wiki!
MRH-Net (for magnetic resonance histology-Net) is a tool developed to transform mouse brain magnetic resonance imaging (MRI) data into maps of specific cellular structures (e.g. axon and myelin).
Modern MRI provides unparalleled tissue contrasts for visualizing brain structures and functions non-invasively. To this date, people are still actively developing new MRI contrasts, or markers, for detecting pathological changes in the brain. However, the sensitivity and specificity of MRI markers to cellular structures, especially under pathological conditions, are largely unknown due to lack of direct links between MRI signals and cellular structures. This is quite different from histology, which, over the past few decades, has developed multiple stainings that tag cellular structures with high specificity. For example, people can use antibodies specifically bind to neurofilament (NF) and myelin basic protein (MBP) to stain axon and myelin, respectively. These stainings are common tools for neurobiologists to study disease mechanisms and potential treatments using mouse models, but the procedures to acquire histological images are invasive and time-consuming. We hope that MRH-Net will assist neurobiologists to take the advantage of modern MRI techniques by providing them easy-to-understand maps of key cellular structures with the highest possible sensitivity and specificity.
MRH-Net are deep convolutional neural networks (CNNs) trained using co-registered histology and MRI data to infer target cellular structures from multi-contrast MRI signals. It was designed for the following tasks:
To transform MRI data to images of cellular structures with contrasts that mimic target histology. To enhance the sensitivity and specificity of MRI to specific cellular structures. Here, you will find our trained MRH-Nets, their source codes, our mouse brain MRI datasets used for the training and testing and the acquisition parameters. The datasets have been carefully registered to mouse brain images from the Allen Mouse Brain Atlas (https://mouse.brain-map.org).
MRH-Net was designed based on three assumptions:
The relationship between MRI signals and target cellular structures is local, so that the signals at each pixel is a realization/instance of the relationship between cellular structures and MRI signals. The millions of pixels in each MRI dataset thus provide sufficient data to train MRH-Nets. The multi-contrast MRI signals are sensitive to the presence of taget cellular structures. Different MRI contrasts are sensitive to distinct aspects of a particular structure (e.g. diffusion MRI is sensitive to restrictive effects of cell membrane and myelin sheath), multi-contrast MRI signals can potentially help MRH-Net and improve the sensitivity and specificity. Deep CNNs can accurately infer the distribution of target cellular structure from multi-contrast MRI signals. While the first assumption is true for most MRI data, it is often difficult to know whether the last two assumptions are met. In our experiments, the MRI protocol includes T2, magnetization transfer (MT), and diffusion MRI, which have been shown to be sensitive to axon and myelin. The results generated by MRH-Net based on multi-contrast MRI data demonstrated remarkable similarities with NF and MBP stained reference histology, and the sensitivity and specificity of the results were higher than any single MRI marker.