• identification and quantification of phenotypes for cell images useful for understanding biological activity in response to drug treatments → traditioanl approach is classifcal image anaylysis, but predictive CNNs with ResNet and VGG16 proves useful → predicting cell mechanisms of action in response to chemical perturbations for two cell profiling datasets from Broad Bioimage Benchmark

  • networks were pretrained on ImageNet, enabling much quicker model training → ability to quickly and accurately distinguish between different cell morphologies from scarce amount of labeled data illustrates combined benefit of transfer learning

  • High-content scrrening helps identify + quantify cell phenotypes → CNNs can help discover features needed for classification of images based on raw pixel intensity data → combination of segmentation and classification in single framework means image clasification can be done without prior cell segmentation

  • major bottleneck of supervised CNNs is scarcity of labeled data → studies have shown reusing models trained on different tasks reduced problems → transfer learning's low-level filters and feature understanding

    • state-of-the-art deep CNNs pretrained on natural images with minimal preprocessing and without segmentation → used for mechanisms of action and nucleus translocation, based only on pixel intensities that automatically pass through network to give final predictions