ASCB EMBO 2018 Poster: Machine learning powered parameter free 2D and 3D image segmentation and obje

Updated: Jul 2, 2019

This poster will be presented by Luciano Lucas, PhD on Sunday December 9th from 12:00 - 1:30 PM. Poster number P1046.

The technology presented in this poster is a part of our latest release, Aivia 8. Request a demo to try it out:


Modern sample preparation, molecular probes and imaging technologies enable researchers to image at high spatial and temporal resolutions. While it is relatively easy to generate large numbers of high quality image data, it is hard to efficiently extract knowledge from them. This is due to a critical limitation in state of the art image analysis tools. These tools use sets of manually engineered algorithms to generate segmentation and analysis results. They require a user to have a solid understanding of image processing algorithms and master several user facing parameters before one can efficiently use the tools. To over come this limitation, we have expanded our machine learning (ML) technology to create a parameter free image segmentation and object analysis pipeline which requires minimal knowledge of image processing algorithms.

In this pipeline, image analysis parameters are autonomously learned from the image data itself. A user only has to define the types of objects he/she is interested in analyzing (e.g. cell cytoplasm and nuclei) by drawing a few regions representing the object types of interest. The tool then learns the intrinsic image analysis parameters to assign object type confidences to each pixel using a customized Random Forest classifier, resulting in full image object confidence maps. This is followed by learning the morphological and intensity parameters needed for object partitioning (i.e. separating touching objects). The user’s drawings are used to infer key descriptors such as size, shape, and separating distance between neighboring objects. This information is used to setup internal parameters to partition objects in the confidence maps and thus creating segmented objects of defined types. This is followed by additional analysis steps such as object measurements and machine learning enabled object classification.The trained algorithm can be applied without any parameter settings to volume images using the batch processing engine of the pipeline and the algorithm scan also be incrementally updated by additional learning.

We applied this pipeline for 2D and 3D microscopy image analyses. The parameter free results are comparable to analysis using human expert set parameters, which validates our approach.

Fast machine learning pixel classification in combination with parameter free image segmentation and object analysis allow users to quickly process large numbers of data with minimal image processing knowledge. Recently, deep learning model is growing in its applicability for image analysis problems but bottlenecked on large numbers of manually annotated training data.The pipeline enables automated annotation to reduce cost and accelerate the adoption of deep learning.

Authors: M.Jones , H.Lai , C.McBride , S.McElroy , J.S.Lee , L.A.Lucas

Authors are a part of DRVISION Technologies LLC, Bellevue, WA