Just Go With The Optical Flow: Enhanced Axon Tracing by Modeling Fiber Connections as Motion

Fin Amin

Abstract:

In this work, we explore the extension of motion-
based techniques from computer vision into the realm of

biological structure detection. Specifically, we propose using
optical flow as a feature for guiding dense axon centerline
detection and tracing. Optical flow is a technique that describes
the apparent movement of objects over time, as in a video. Our
experiments show that interpreting 3D volumes as videos, by
interchanging the time axis of optical flow with the x, y, and
z axes of the volumes, extracts useful spatial relationships in
the context of axon centerline detection. This approach offers
several advantages. First, optical flow captures subtle spatial
gradients, providing a rich representation of directional changes
in intensity, which helps delineate the linear structure of axons.
Second, by leveraging the continuity of flow vectors, our method
enhances the detection of axon trajectories even in areas with
low contrast or noise. Our experiments demonstrate that this
adaptation of optical flow yields superior results in centerline
detection compared to traditional edge-based or segmentation
methods. To validate our claims, we show performance on
two axon tracing datasets. By interpreting optical flow in this
multidimensional space, we offer a new perspective on how
spatial dynamics can inform 3D tracing tasks.