Automated Analysis of Microscopy Image Based Drug Screens with Deep Multiple Instance Learning
High-content screening (HCS) technologies have enabled large-scale microscopy imaging experiments for studying cell biology and for drug screening. These systems produce hundreds of thousands of microscopy images per day and their utility depends on automated image analysis. Our team developed deep learning approaches that learn feature representations directly from pixel intensity values, rather than existing analysis techniques that rely on segmentation and feature extraction based pipelines. As most deep learning pipelines typically relied on having a single centered object per image, these methods were not directly applicable to microscopy datasets. We developed a segmentation free approach that combines deep convolutional neural networks (CNNs) with multiple instance learning (MIL) in order to classify and segment microscopy images using only whole image (i.e. treatment or well) level annotations. Since developing this method we’ve built a scalable platform around it and successfully applied the method to numerous assays including proteome-wide genetic screens in yeast, and drug screening and profiling assays in both fibroblast and cancer cell lines. We’ve also built a secondary weakly-supervised workflow, that allows researchers to cluster and visualize treatments in HCS screens by how similar they are phenotypically. In the presentation, we’ll introduce these more recent results and describe how the two workflows, enabled by our platform, can be used to analyze almost any HCS screen automatically.