A practical guide for both 2D (satellite imagery) and 3D (medical scans) image segmentation using convolutional neural networks. A tour through a complete Jupyter notebook - data preprocessing (OpenCV/SimpleITK), neural network implementation (Keras with TensorFlow backend) and eye pleasant visualizations in the open-source K3D Jupyter module. All topped by hints and cases from real projects.
Machine learning in the world of image processing is developing faster than ever before with convolutional neural networks (CNNs) as the most common tool. As researchers and developers, we still look for methods to keep up with progress in the variety of network architectures, layers and loss functions. Jupyter notebooks seem to be the perfect environment for fast prototyping, debugging and sharing results using the power of Python and its rich diversity of modules. The goal of this talk is to show how our team benefits from these in real projects. On the example of image segmentation task, we would like to present our road to current solutions and algorithms. The example is going to be focused on a practical guide including: - a really short introduction to 2D/3D image segmentation and medical image analysis with the SimpleITK module - a review and Keras implementation of CNN architectures used for image segmentation - a presentation of results using the open-source K3D Jupyter module for which members of our team greatly contributed to All steps are going to be filled with hints from experience to help participants with a quick start in the domain and prevent them from repeating some popular mistakes