View All Posters

P2071
Automatic Recptor Trafficking Assay Analysis using Machine Learning
Presenter James Lee, SVision LLC, USA
Additional Authors: Samuel V. Alworth, Bryan Ho, Louis Piloco, Anthone W. Dunah, Greg V. Martin SVision LLC
The spatial temporal dynamics of surface receptors is a common phenotype of interest in neuroscience imaging assays, and is important to inter cellular communication and brain function. The assay result depends on the quantitative characterization of small receptor clusters just a few pixels in diameter. The subcellular phenotype often exhibits weak and unstable signal that is sensitive to noise and variations. This imposes a critical limitation on the image based assay outcomes.
We have previously validated that a subcellular object detection method using confidence maps could enable highly accurate and robust quantitative analyses of the synaptic vesicle recycling assay. We extended the approach for broad applications by supplementing the validated approach with a learning module. This allows the user to modify and optimize the approach automatically for different applications using a drawing tool on images directly. We previously validated that the learning module could achieve the same high level of accuracy and robust performance as the validated method using images from a synaptic vesicle recycling assay as well as the standard Transfluor high-content screening assay.
We further validate the effectiveness of the learning module using neurotransmitter trafficking image sets provided by our collaborators at the MassGeneral Institute for Neurodegenerative Disease. Our hypothesis is that the learning module will have similar results to that of a commercially available benchmark. We used detection sensitivity, positive predictive value, detection (segmentation) accuracy, and assay quality metrics as the performance metrics. The image truth is created manually and validated by independent review and update. We reject the hypothesis with statistical significance. We conclude that the learning module, which can be programmed through drawing, achieves significantly better performance than the commercial benchmark.