What is the structure of the neural circuitry of the brain? How does it relate to cognition, aging, and disease? These are no longer questions to be answered by neuroscientists alone. Scientific research is increasingly tool driven, and computer scientists play an essential role in both developing and evaluating the tools used for making scientific progress. This was true in the earliest days of the computer with the championing of computational simulation by von Neumann, and it remains true today with the modern computational science advocated by Jim Gray, which includes large scale collection, management, automated analysis, and visual exploration of scientific data. In a more personal sense, Fred Brooks also described one role of the computer scientist as that of a toolsmith who uses interdisciplinary collaboration not only to make advancements in driving problems of collaborators but also to do better computer science research. For both the individual computer scientist and the field as a whole, there is much to gain from exploring and expanding this interface with the many scientific subfields.
Of course, there is a huge landscape covered by this kind of collaborative scientific research – while working on my PhD at Brown CS I have been exploring the area at the intersection of brain imaging and computer science. This area is particularly exciting due to the ever-increasing quality and quantity of in-vivo imaging data made available through magnetic resonance imaging (MRI). This kind of imaging is widely used to help us understand the brain; however, the complexity of both the data and the underlying anatomy makes computational tools crucial in any application of MRI.
My research specifically investigates tools for understanding brain white matter using diffusion MRI, a technique that is sensitive to the geometric properties of tissue microstructure. In combination with computational tools, diffusion MRI is useful for reconstructing a large-scale wiring diagram of the brain that represents the three-dimensional structure of white matter. These white matter connections constitute an array of communication channels between brain areas, and the healthy functioning of the brain depends greatly on their structural integrity. Image-based reconstructions are useful for visualizing this anatomy in specific individuals, but they are also useful for quantitative image analysis and the characterization of anatomical variation across populations. While there has been much successful work to develop these methods over the years, many open problems persist in scalable image analysis of large populations and in computational modeling of complex features of anatomy.
My research has focused on developing and evaluating tools to address these challenges in three areas: voxel-based analysis, fiber bundle modeling, and multi-fiber image processing. The first area includes the most common way to analyze diffusion MRI datasets, in which individual voxels or regions-of-interest (ROIs) are compared across a population. We’ve shown how ROIs can be automatically defined without expert annotation, drawing inspiration from supervoxel segmentation algorithms from computer vision. We have also done extensive evaluation of the variety of methods used in common practice, which can help scientists better understand how different pipelines compare in terms of reliability and sensitivity.
The second area examines methods for fiber bundle reconstruction, which provides geometric models that are more analogous to what anatomists have historically studied through dissection. Our work in this area has developed a novel representation of curve data with a sparse closest point transform. Our experiments have shown how this enables a wider range of tools in statistics and machine learning to be used to achieve scalable and accurate performance in common fiber bundle modeling tasks, such as clustering, simplification, and population-based selection.
The third research area looks at image processing with diffusion MR models consisting of multiple fibers per voxel, which has potential to improve the quality of anatomical reconstructions in both individuals and population averaged data. For this purpose, we extended a kernel regression framework for image processing to operate on diffusion model data, and our experiments show how this is useful for interpolation in tractography and data fusion in atlas construction. One practical outcome of this has been fiber bundle models from a population average of 80 people that demonstrate more complete features than typically found in diffusion MRI atlases. We have also applied this to fiber bundle modeling in the presence of brain tumors, demonstrating potential applications to clinical imaging. At a high level, these three areas represent somewhat overlapping ways to analyze diffusion MR imaging data, but one lesson learned in studying them is that there is no best method. Instead, they form a toolbox with many possible uses, and we can benefit from evaluating these tools to better understand how they relate to the data and the scientific question at hand in any particular application.
Looking forward, these tools have potential to help improve our scientific understanding of the brain through ongoing population imaging studies of normal variation and neurological and neuropsychiatric diseases. Specifically, our group has ongoing collaborations with neuroscientists and clinicians investigating normal aging, cognition, HIV, Alzheimer’s disease, and pediatric bipolar. These collaborations have been important for developing and evaluating our methodology, but we also aim to extend our scientific knowledge of these processes. There are also potential applications of some of this work to surgical planning, guidance, and validation, e.g. for improved electrode placement in deep brain stimulation and for more accurate planning of tumor resections.
Beyond the specific applications, there are also many remaining open problems related to the methodology. One challenge is to “zoom out” to better understand large populations, and many of these challenges here are computational and relate to data management, exploration, and scalable analysis. We also have much to gain by “zooming in” to better understand anatomy at increasing fine anatomical scales, possibly by jointly visualizing anatomy with both MRI and microscopy and also by improving analytical tools to make best use of improvements in imaging hardware. While we do not yet know what we will discover about the brain in the process, both interdisciplinary collaboration and computational tools will surely be important for gaining that knowledge.