Training

After this course you should be able to:

• Explain the fundamentals of machine learning methods suitable for image analysis

• Consult users/colleagues in strategies to obtain ground truth

• Give advice in training and using a neural-network

• Perform simple quality control on the results of one selected ML approach

Full details for the practical sessions (participants choose to join one group):

Detection and Segmentation in microscopic images (Thorsten Falk, University of Freiburg)

Neural Networks are extremely powerful in learning various image processing tasks. Their biggest strength is their easy adaptability to very different applications by simply training them with application-specific corresponding observation-outcome pairs. Especially in biology where imaging is by no means standardized and the number of applications on various scales is uncountable, neural networks can be the swiss army knife for image analysis. However, the required observation-outcome-pairs for training are also their biggest weakness. The key for good models is a sufficient amount of high quality training data covering the distribution of possible observations.

The focus of this workshop will be instance detection and segmentation in biomedical image data which can be solved with a comparably simple convolutional-deconvolutional feed-forward network. Image segmentation is one of the most frequently required but also one of the most demanding tasks with respect to training data generation. Each pixel of multiple images must be accurately and consistently annotated to obtain training data for a neural network. You will learn how to properly record, select and annotate images to train a neural network for your task. You will employ data augmentation techniques to reduce the image annotation effort to a moderate level. Finally, the concept of domain adaptation to transfer learnt models from one imaging device to another will be explained in a nutshell.

The workshop will require but also teach manual annotation in Fiji/ImageJ and use a new Deep Learning plugin for simple 2D tasks. 3D tasks can be tackled in a similar fashion but will require python and the caffe framework (this may change to keras/tensorflow depending on preparation time). Attendees need to provide 2D or 3D microscopic data for structure detection and/or segmentation with corresponding point markers indicating the positions and types of structures to be detected or (potentially multi-class) segmentation masks for training networks for structure segmentation.

Segmentation in 3D microscopy image stacks (Anna Kreshuk and Constantin Pape, EMBL)

We will focus on segmenting 3D data with convolutional neural networks. We’ll talk about network architectures that were shown to work well in 2.5D and 3D, about pre-processing and data augmentation, as well as necessary post-processing. We will also introduce and compare tools for groundtruth annotation and proof-reading in 3D.

Transfer learning and how to use synthetic data for supervised deep learning (David Rousseau, Pejman Rasti, Université d’Angers)

"A striking fact when gazing at the first layers of deep neural networks is that these layers almost look like Gabor wavelets. While promoting a universal framework, these machines seem to systematically converge toward tools that humans have been studying for decades. This empirical fact is used by computer scientists in the so-called transfer learning where the first layers of an already trained network are re-used to save time and improve results when limited data sets are available. Another current limitation to the application of deep learning is that this requires huge training data sets to avoid overfitting. Such training data sets have to be annotated when supervised learning is targeted. And, at the moment, the available data sets available among the bioimage analysis community are far from covering the large amount problematics in bioimaging. Here also some possible alternatives exists. This includes the generation of realistic automatic annotated data sets with synthetic simulated images.

During this hands on you will learn the basics of transfer learning with a Jupiter notebook and will learn how to use synthetic data for supervised deep learning. This can be applied to any type of bioimaging problem. A large panel of examples will be given for illustration among which 2D microtubule imaged with PALM/STORM, 3D cells in spheroid imaged in light sheet fluorescence microscopy and 3D plant roots images in absorption X-ray tomography."

Content-aware image restoration (Martin Weigert, MPI CPG)

Fluorescence microscopy of living cells or organisms often results in noisy and axially blurred images, owing to the limited laser power and number of focal planes that are compatible with specimen health or temporal resolution. This is in contrast to acquisitions of fixed specimen, where such constraints are often not present, and where the signal-noise-ratio (SNR) and axial sampling can be chosen almost freely.

In this part of the tutorial, we will explain how machine learning can be used to improve the quality of images that were acquired under adversarial imaging conditions, with the help of adequately acquired high quality images. Participants will learn how to construct image restoration pipelines that are tailored to specific imaging situations and organisms, and will be guided through all the necessary steps: From data pre-processing, model training, and application to new images to the final deployment via Fiji plugins that can be shared with collaborators.

Ideally, the participants should already have a specific experimental situation in mind, where normal imaging conditions would preclude sufficient image quality. To take full advantage of this practical session, they are encouraged to bring already acquired corresponding images/volumes of samples acquired at different (low and high quality) conditions.