FIGSS Seminar

Since the 1970s computer assisted analysis of microscopy data has been used, but even nowadays a high level of human expertise and input is necessary to analyse the data. Modern developments in microscopy technology have made it possible to produce large amounts of data from a single experiment. The analysis of such data has therefore become a bottleneck in experimental procedures. We are developing data analysis methods that are based on generative models to help relieve the experimentalists from this bottleneck. Generative approaches attempt to understand the image by creating an internal representation of the data. Even if a full understanding cannot be reached it can greatly simplify the work of an experimentalist by limiting the amount of work that has to be performed by hand. We illustrate the usefulness of generative models by combining them with well developed segmentation processes in the task of finding and counting circulating tumour cells (CTCs) obtained by fluorescence microscopy. Counting CTCs in the blood is an important diagnostic tool in various types of metastatic cancer. A high count indicates that the current treatment is not effective and rapid progression of the disease is imminent. We use the so-called expectation maximisation (EM) to learn features of the CTCs. The learning of features with EM, rather than defining them by hand, makes it easy to adapt the framework to similar tasks.

Transcranial magnetic stimulation (TMS) allows to non-invasively manipulate neural activity via strong magnetic fields. TMS is currently tested for its clinical use for treating depression, stroke, schizophrenia and several other neurological disorders. However, the details of how TMS induces patterns of neural activity in cortical circuits remain poorly understood, which hampers targeted clinical application. Assessing the nature of such patterns or establishing the biophysical basis underlying magnetic stimulation in purely experimental settings remains difficult given the scarce recording opportunities and high variability of results across healthy subjects. Models incorporating anatomically detailed neurons could overcome these limitations and provide valuable insight into the effects of TMS at the cellular level. We develop a computational model of how TMS stimulates cortical circuits and explain how TMS produces I-waves, fast rhythmic responses in descending motor pathways. The model explains findings from a range of experiments with different stimulation protocols and is an important first step towards designing optimized stimulation protocols for specific clinical purposes.

Retinal ganglion cells (RGCs) are nerve cells in the retina responsible for transmitting visual signals from the eye to the brain. It is known that RGCs at different locations of the retina have different properties regarding their encoding of orientation, frequency, size, shape, etc. [Croner and Kaplan 1995; Schall and Leventhal 1987]. Similarly, many studies show that features like contrast and orientation are not equally discriminated across the visual field, i.e., humans have perceptual biases across retinal position [Helley et al 1996, Rovan et al 1982, Berkley et al 1974]. However, the reasons for these position dependences of RGC properties are not well understood. Classic theories of sensory coding in the brain posit that sensory systems take the statistical regularities present in the environment into account in order to reduce redundancies in their representation. To model these environmental regularities, we construct forests in a virtual environment and extract images from the viewpoint of a human agent walking through them. To mimic the natural input to the visual system, these images are projected onto a spherical eye ball model by a generalization of the thin lens model [Hecht 2001; Helmholtz 1855]. Finally, we quantify the second order dependencies of patches from the spherical images taken at various retinal positions and estimate the corresponding optimal whitening filters [Dong and Attick 1995], i.e. those filters that optimally reduce the redundancies (in terms of correlations) of the input. We find that these filters have properties similar to retinal ganglion cells' receptive fields. Our results suggest that the encoding of visual information by the retina is an adaptation to both the statistics of the natural environment and the imaging geometry and shape of the eye.

At the moment, we see a drastic change in the European energy supply system. Fossil fuel use is reduced to lower CO_2 emission. Nuclear power is viewed skeptically in many countries, and a phase-out is already scheduled or even executed in countries such as Belgium (2025), Germany (2022), Italy (1986) and Switzerland (2034). In our research, we address the question how the future European electricity supply may look like. If the EU 2050 targets are to be met, electricity must come almost entirely from CO_2-free generation by then, with a decreasing share of nuclear power. A substantial fraction will have to be covered by the weather-dependent sources wind and solar power. This completely alters the way we have to look at our electricity system. While conventional plants prevailing today are all dispatchable and can be regulated to meet the demand, weather-dependent sources have intrinsic fluctuations to which we will have to adapt on demand-side, be it with storage, smart grids, increased European transmission or backup hydro power. The effects of such adaptions and how they could be combined to keep the (huge) necessary effort feasible is the topic of this talk.