Extracting 3D brain surface from MRI
Problem
Computing brain volume and shape are a key diagnosis information used for many neurodegenerative diseases such as Alzheimer’s. 3D MRI provides exquisite tissue definition with a typical resolution of 1mm3, but is difficult to analyse quantitatively by human expert. State-of-the-art image analysis software can take hours of processing to extract the brain surfaces from a single scan, challenging clinical adoption.
We have setup a large collaborative team to use novel AI method (deep learning) to compute brain surfaces from a 3D brain MRI. Our goal is to reduce the processing to a few minutes without compromising accuracy. The project is a large collaboration between Maxwell+, QUT, CSIRO Data61 and the Australian eHealth Research Centre. The project was funded in part by a CRC-Project grant.
Computing the brain surface from a 3D MRI is not a trivial task, fraught with several challenges: brains have highly complex shape and vary considerably between individuals and during the course of diseases, the MRI resolution is not high enough to resolve all the fine details of the anatomy resulting in partial volume effect, and the estimated brain surface needs to be free of artefacts to allow enable clinical diagnosis.
Our solution
In one approach we designed a deep learning model that can learn to generate a surface directly from a 3D scan (DeepCSR method). It creates a map at any arbitrary resolution of the interior and exterior of the brain, which can then be further processed to compute surfaces. More details in our paper that won the best paper award at WACV21.
Brain surfaces can be computed in less than 30min at different resolution, with further improvement planned:
For more information, contact us.