Research Highlights

Efficient, Scalable, and Cost-effective 3D Microscopy Image Segmentation Using a Microlabor Workforce

box
3D rendering showing two axons in red and blue.

January 2014 La Jolla

January, 2014 La Jolla -- In an applications note published in Bioinformatics, a team of researchers from the National Center for Microscopy and Imaging Research at UCSD described a new scalable, semi-automated approach to segment 3D structures revealed in serial block-face scanning electron microscopy (SBEM) images. The team used the Dual Point Decision Process (DP2) to divide the segmentation problem into discrete tasks distributed to a large pool of individual workers.

Amazon’s Mechanical Turk system provided access to the micro-labor workforce. SBEM techniques, coupled with new staining protocols, can reveal cell boundaries, sites such as synapses, and many intracellular components, such as synaptic vesicles and mitochondria. The improved resolution and amount of detail afforded by SBEM, among other techniques, is enabling researchers to realistically explore new scientific questions pertaining to morphology and network connectivity.

In cellular imaging, though, manual segmentation is a well-recognized bottleneck. In a typical scenario, it involves a single trained expert using automated algorithms or manual methods to examine and mark up each individual slice and trace contours around the structures of interest using a program such as TrakEM2 or other specialized software programs. DP2, however, streamlines and parallelizes the process by distributing it to a large number of workers available through Amazon’s Mechanical Turk system.

This approach differs from other automatic segmentation techniques because, rather than relying on trained users, it uses a simple web interface to reach a large number of unskilled workers who are asked to make binary decisions in response to simple yes-or-no questions related to whether two points in an image were placed on the same object. This approach can scale to the number of workers available for assignment.

In tests, tens of thousands of decisions were accomplished in 51 days. The average cost was $1.2 (USD) per cubic micron for data set 1, a mouse optic nerve sample, and $56 per cubic micron for data set 2, a mouse cerebellar neuropil sample. Each decision received a payment of one cent.


Citation: Richard J. Giuly, Keun-Young Kim, and Mark H. Ellisman, DP2: Distributed 3D Image Segmentation Using Micro-labor Workforce. Bioinformatics Applications, 2013, 29: 1359-1360, doi:10.1093/bioinformatics/btt154.

Link to Online Article

Python-based code for non-commercial use and test data