Title: Automated image analysis of prion proteins with deep learning
Speaker: Thomas de Mondesir
Abstract:The objective of this talk is to introduce the development of a robust analysis pipeline to analyseAFM (microscope) images. These pictures are essential for the understanding of the mechanisms behind prion protein aggregation formation. To cope with our lack of training data, we introduced an image generation method that can produce training examples (X,Y) (X is an image, Y is either a segmentation mask or a measure) for supervised deep learning models. We train the Mask R-CNN instance segmentation model on our generated data to learn how to extract every oligomer from an AFM image. The next step is to collect information at a scale where an observation is an aggregate. We create a simple neural network for regression, which after training on artificial images can measure the length of an aggregate. To study the composition of oligomers we use Mask R-CNN once again, but in a different setting, the goal being to detect spherical elements called S. Our approach is the sequence of the mentioned steps, a pipeline that takes an image of oligomers as an input and determine the location, length, surface area, and the number of S elements of every aggregate.
** This work has been achieved during my Master‘s internship from April to August 2020.