Discovery of Rare Phenotypes in Cellular Images Using Weakly Supervised Deep Learning

Heba Sailem, Mar Arias-Garcia, Chris Bakal, Andrew Zisserman, Jens Rittscher; Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 49-55

High-throughput microscopy generates a massive amount of images that enables the identification of biological phenotypes resulting from thousands of different genetic or pharmacological perturbations. However, the size of the datasets generated by these studies makes it almost impossible to provide detailed image annotations, e.g. by object bounding box. Furthermore, the variability in cellular responses often results in weak phenotypes that only manifest in a subpopulation of cells. To overcome the burden of providing object-level annotations we propose a deep learning approach that can detect the presence or absence of rare cellular phenotypes from weak annotations. Although, no localization information is provided we demonstrate that our Weakly Supervised Convolutional Neural Network (WSCNN) can reliably estimate the location of the identified rare events. Results on synthetic data set and a data set containing genetically perturbed cells demonstrate the power of our proposed approach.