Our Blog
09 MaySBMT2013 Poster
An ever-improving model of the structure of the living human brain
Andrew Worth1, Jason Tourville2
1 Neuromorphometrics, Inc., Somerville, MA
2 Dept. of Speech, Language, and Hearing Sciences, Boston University, Boston, MA
In a traditional brain atlas the anatomist aggregates experience of many individuals and presents it as a single brain. This establishes what is normal, but individuals differ from each other, to a greater or lesser extent, in each neuroanatomical region. We are building a model of brain structure by labeling each voxel in a large number of individual MRI scans. Our goal is to continuously improve the model, not only by adding more scans, increasing the number of anatomical regions, and by parcellating regions into more specific sub-regions, but also by iteratively reviewing the results in the context of the whole model to find and fix errors.
Magnetic resonance can provide only an imperfect representation of the highly detailed “true” neuroanatomy. Beyond noise and artifacts, MR signals have only macroscopic resolution and mainly indicate myelin content. But because MRI is non-invasive, it is clinically relevant: we are building a model of the living human brain. And the significance of this model is that it provides quantitative information about normal variation as a result of individually labeling a large number of scans.
The model was created by specifying neuroanatomical boundaries and extents and using in-house software, “NVM” to comprehensively label each brain voxel in every scan. Raw MRI scans were preprocessed (for example, positional and intensity normalization), and labeled scans were added to a probabilistic atlas and normative statistics on volume, location, shape, and intensities were compiled into a database. All structures in each individual scan were then compared to the atlas and statistics in a “bootstrapping” process to check, correct, and refine the boundaries and voxel labels.
Protocols defining 157 ROIs were used to manually label 64 T1-weighted MRI scans. A probabilistic atlas and normative statistics were generated to refine the labeling and then the atlas and statistics were re-created. We show example structures from the atlas and give examples of the neuroanatomical variation by showing typically different gyral and sulcal configurations.
Even beyond the multi-decade effort needed to develop the system and train technicians, the amount of work necessary to get to this point was enormous. The total technician time needed was more than 2 years. The initial time required to label every structure in every slice of a single scan started out at about a week and decreased to 2-3 days. While ongoing improvements in automation will lower the cost, it is still high enough to be unjustifiable. The only way to increase the number of labeled scans is to spread out both the funding for the work and also the use of the results. We provide academic researchers with access to all scans in the database for less than the cost of adding a single scan to that database. As multiple researchers contribute, the database can continue to grow.
Releasing the model will help in creating automation and provide a database to mine, which will lead to the discovery of biomarkers and to an understanding of the variation of brain structure.