Real-time Alignment for Connectomics

Neha Goyal, Yahiya Hussain, Gianna Guang Yang, Daniel Haehn

Keywords: image alignment, registration, feature matching

TL;DR: We hypothesize that adding biological features to register the images results in faster alignment.

Abstract: In Connectomics, researchers are creating the brain's wiring diagram at nanometer resolution. As part of this processing workflow, 2D electron microscopy (EM) images must be aligned to 3D volumes. However, existing alignment methods are computationally expensive and can take a long time. We hypothesize that adding biological features improve and accelerate the alignment procedure. Since especially mitochondria can be detected accurately and fast, we propose a new alignment method, MITO, that uses these structures as landmark points. With MITO, we can decrease the alignment time by 27\%, and our experiments indicate a throughput of 33 Megapixels/sec, which is faster than the acquisition speed of current microscopes. We can align an image volume of $1268\times1524\times160$ voxels in less than 12 seconds. We compare our method to the following feature generators: ORB, BRISK, FAST, and FREAK.