shz9 / deeploc-vae

Semi-supervised VAE model for protein localization prediction from microscopy images

DeepLoc-VAE

Semi-supervised Classification of Protein Subcellular Localization Patterns in Microscopy Images with Variational Autoencoders

Author: Shadi Zabad, University of Toronto

This is a report on my course project for CSC2548 (Machine Learning for Computer Vision), taught by Sanja Fidler.

Recent years have witnessed an unprecedented growth in the number and scale ofmicroscopy images generated from a variety of experimental protocols. In response,numerous image analysis pipelines have been proposed to automate various aspectsof the process of scientific discovery, from image segmentation to high-levelannotations. One particular research area that has garnered some attention is theclassification of protein subceullar localization patterns from microscopy images.Previous methods applied in this setting relied on fully-supervised machine learningtechniques to automate the annotation process. Here, we improve on previous workby deploying deep generative models to perform semi-supervised classificationof protein subcellular localization patterns. We show that, despite using only asmall fraction of the labels, the semi-supervised generative model achieves state-of-the-art results. In addition, we leverage the generative capabilities of the modelto explore some of the properties of the inferred latent representations.

Qualitative Assessment of the Generative Capabilities: