Experience is the best teacher: Common pit-falls and misconceptions about machine learning based 3D image segmentation

Electron – Scientific Presentation 14:15 – 15:25 (Sydney Time) | Wednesday 17 Feb 2020


Large volume 3D EM imaging promises to revolutionise several different fields of biology, but the unwieldy quantities of data generated by these techniques are a obstacle to their application. Meaningful analysis of volumetric EM data traditionally relies on the manual segmentation and classification of interesting structures by a skilled microscopist (approximately 2 hours of labour per cubic micron); A few thousand cubic microns typically requires several months of full-time work. Machine learning is a sub-field of artificial intelligence with a strong potential to streamline the 3D image processing, by automating the most laborious aspects of image segmentation and classification. Unfortunately, despite the clear value of these capabilities, microscopists often find their first foray into machine learning frustrating, underwhelming, and highly discouraging. I will outline simple solutions to the pit-falls that typically dissuade microscopists from using machine learning based computer vision. I will also outline the fundamental principles of a new processing strategy called AI-directed Voxel Extraction (AIVE), which resolves one of the most under-appreciated limitations of machine learning to enable the processing of >3000 cubic microns in just a matter of days.


Benjamin Padman

Benjamin Padman

The University of Western Australia

Bio available soon