The six main tasks in image processing: an overview

• Philip Kollmannsberger (Heinrich-Heine-Universität Düsseldorf)

Modern imaging methods enable us to capture structure and dynamics in unprecedented detail and with high temporal and spatial resolution. The challenge is to make full use of the potential of such "big" imaging data to obtain quantitative results, test hypotheses, and develop new theories and models. Manual or semi-automated image analysis workflows quickly become a bottleneck because they do not scale well with the amount and complexity of the data. This lecture provides an overview of typical automated image analysis workflows using examples from high-resolution microscopy in the life sciences. We will cover the six main tasks, including image reconstruction from raw tomography or localization data, denoising, tracking in time and space, segmentation to extract objects from images, visualization, and explainable AI-based methods. In the last part, we will introduce tools for integrating these different tasks into complete workflows.

Tomographic methods in medical imaging

• Christoph Lerche (FZJ)

In medical imaging, dedicated cameras are used to capture images, and medical images are typically three-dimensional images. These three-dimensional images are generated from a series of datagrams or images of sections and/or projections, which is achieved by tomographic reconstruction. Magnetic Resonance Imaging (MRI), Computed Tomography (CT), Positron Emission Tomography (PET) and Single Photon Emission Computed Tomography (SPECT) are typical examples of different imaging modalities where tomographic reconstruction methods are used. This lecture gives a brief introduction to Fourier transform based reconstruction (MRI), filtered back projection (CT) and iterative reconstruction (PET and CT).

Visualizing spatial datasets

• Deborah Schmidt (MDC)

Join us in our journey through selected methods of telling the story of your dataset through visualizations. In the seminar, we will transform 3D segmentations into effectful Blender renderings.

Tracking of objects: from one to many

• Carsten Rother (Heidelberg University)

I will talk about two separate research areas for tracking objects over time and instances. The first applies to the scenario of tracking an object over time, e.g. a known 3D rigid model. I will introduce Bayesian and particle filters and explain some technical ideas to make them fast and accurate. In the second part of the lecture I will introduce the field of tracking a large set of objects (e.g. cells) over time, and also instances. Since the objects typically do not move randomly, it makes sense to formalize their "structured motion" and formulate this task as a structured prediction problem. To this end, I will introduce efficient solvers for this problem.

Explainable Machine Learning

• Ullrich Koethe (Heidelberg University)

Explainable machine learning involves two complementary types of questions: (1) Method designers and theorists ask: "Why do deep neural networks work so well? How can we further improve them? How can we provide formal performance guarantees?" (2) Instead, method users ask, "How did the network arrive at its conclusion? What variables determine the result, and in what way? Can the result be trusted?" The talk will introduce both areas and review the current state of our answers.

Image Denoising: Supervised, Self-supervised, Generative

• Alexander Krull (University of Birmingham)

Removing imaging noise is an essential problem in scientific applications, where sensors are often pushed to the edge of what is possible. The past years have seen a range of machine learning methods proposed: supervised approaches, using image pairs to learning a mapping from noisy to clean images; Self-supervised approaches, capable of learning such a mapping from noisy data alone; and finally generative approaches capable of additionally capturing the inherent uncertainty of the problem. In this talk will talk about how these approaches can be understood and derived from a probabilistic perspective.

Machine Learning for Image Segmentation

• Dagmar Kainmüller (MDC Berlin)

The course will introduce current machine learning models for image segmentation, including semantic- and instance segmentation. It will also cover how to apply such models to large data. A respective tutorial at the end of the seminar series will give hands-on experience in training and applying such models.