Networks of curvilinear structures are pervasive both in nature and man-made systems. They appear at all possible scales, ranging from nanometers in Electron Microscopy image stacks of neurons to meters in aerial images of road and river networks, and even to petameters in dark-matter arbors binding massive galaxy clusters. Modern imaging systems can produce vast amounts of data featuring them. However, in spite of many years of sustained effort, fully automated exploitation remains elusive, especially when the images are noisy and the linear structures exhibit a complex morphology. As a result, exploiting the image data still requires extensive manual intervention that is time-consuming, tedious, and expensive.
Under ERC funding, we have developed automated delineation algorithms that outperformed the then-current state-of-the-art. Given a 3D stack of microscopy images featuring intricate networks of dendrites and axons, they build tree-like structures that model the connections between neurons. Building such representations on a large scale is key to understanding how the brain is wired and our technology has the potential to enormously speed up this building process. Among other things, it will allow neuroscientists to handle much larger volumes than they currently can. It will also boost their ability to study large groups of neurons, to model their behavior and function, to simulate and classify brain cells, and ultimately to better understand how the human brain works.
However, this will only happen if we can transfer our algorithms from our Computer Science lab into the hands of practitioners for use in their daily work. This is the challenge we have addressed in this project by developing interfaces and tools that make our algorithms both easier to use and more effective.
