Deep learning has demonstrated astonishing classification and segmentation results in microscopy, outperforming all existing approaches. While many codes are publicly available, they require expertise that most biologists lack. The goal of this workshop is to learn how to train and process deep convolutional neural networks for image classification and image segmentation. More specifically, participants will learn how to install python packages and run Jupyter notebooks, train deep learning classifiers to classify images with or without transfer learning, use the ImageJ plugin Annotater to manually annotate images, train deep learning approaches and use them to segment tissue and nuclei, in 2D and 3D images. This workshop does not require proficiency in any coding language.
This is a 3-day workshop. Slides describing the background required to understand the methods and links to video tutorials are available in Courses. All the codes and data used during the workshop are located in Codes and Data.
Software installation
Colab setup
Image classification - MNIST training
Image classification - MNIST processing
Image classification - MobileNetV2 training with ImageNet transfer learning
Image classification - MobileNetV2 processing
Tissue segmentation - Data preparation
Tissue segmentation - UNet training
Tissue segmentation - UNet processing with a Jupyter notebook
Tissue segmentation - UNet processing with Fiji
Nuclei segmentation - UNet training
Nuclei segmentation - UNet processing (Jupyter notebook and Fiji)
Nuclei segmentation - Stardist training (with and without transfer learning)
Nuclei segmentation - Stardist processing (Jupyter notebook and Fiji)
Please cite our papers if you use our codes:
Thierry Pécot, Alexander Alekseyenko and Kristin Wallace (2022): A deep learning segmentation strategy that minimizes the amount of manually annotated images
Thierry Pécot, Maria C. Cuitiño, Roger H. Johnson, Cynthia Timmers, Gustavo Leone (2022): Deep learning tools and modeling to estimate the temporal expression of cell cycle proteins from 2D still images