• Admin

ASCB-EMBO 2019 Poster: GPU-accelerated Machine Learning-powered 3D Image Segmentation at Scale

Updated: Dec 6, 2019

Presentation time: Sunday Dec 8th at 1:30PM

Poster number: P57/B58

The technology presented in this poster is a part of our latest release, Aivia 8.8. Request a demo to try it out: https://www.drvtechnologies.com/demo


Despite the recent advances in bio‐imaging, the extraction of meaningful insights from large multi‐dimensional microscopy images continues to be a major bottleneck to the advancement of science. Non‐machine learning tools use manually engineered algorithms to generate results. They are inflexible and require a user to master several user‐facing parameters before one can efficiently use the tools. By contrast, machine learning (ML) based solutions, only require the user to draw a few regions representing the object types of interest, are starting to be adopted by researchers as they are much easier to master [1]. However, ML based image processing approaches suffer from a different drawback, they are computationally expensive and thus benefit from processing optimizations so that they can be used at scale. Here we discuss enhancements made to our ML‐based image segmentation solution, Aivia’s Pixel Classifier. We have implemented a wide range of 3D filters optimized for 3D data. Moreover, we have implemented a new train and apply pipeline which uses GPU, instead of CPU processing. The new GPU‐accelerated ML‐based image segmentation approach was compared to our previous state of the art ML‐based image segmentation framework presented at this conference in 20181. We used a test image with the following (xyz) dimensions: 607x531x61, 32 bit. Computer environment used: Windows 10, 16 GB RAM, NVIDIA GeForce GTX 1060 GPU, i7‐6700HQ CPU @ 2.6 GHz. In two class segmentation, the GPU accelerated solution achieves 13 times speed up (55 seconds vs. 4 seconds). Moreover, RAM usage was reduced by 35x (7 GB vs. 0.2 GB). Next, we will benchmark the new approach using multiple image sizes as well as >2 segmentation classes. In conclusion, we found that the new solution was ~13x faster while requiring 35x less memory (RAM). This will allow researchers to benefit from the ease of use of ML‐based image segmentation approaches while being able to process large data sets and/or multiple data sets.

[1] M. Jones, et al. (2018). Machine learning powered parameter free 2D and 3D image segmentation and object analysis pipeline. Mol Biol Cell 28, 3727 (Abstract: P1046). https://www.molbiolcell.org/doi/suppl/10.1091/mbc.E18-10-0647/suppl_file/2018ascb-embomeeting-posterabstractsfinal.pdf


M. Jones, C. McBride, T. Lancon, Q. Tran, H. Lai, S. McElroy, S. J. Lee, L. A. G. Lucas

  • All authors are part of DRVISION Technologies LLC, Bellevue WA