- Charless Fowlkes – Learning to detect, segment, and classify biological structures
- Fine-grained classification, pixel-accurate labeling and instance segmentation pose difficulties for standard deep CNN models, particularly when training data is limited. I will discuss our recent progress in designing output domain-adapted architectures that address these challenges and discuss their use in several bio-imaging applications including: meso-scale tracing of neuronal circuits, identifying molecular determinants of animal lifespan, and estimating plant biodiversity from the fossil record.
- Pascal Fua – Comparing synaptic structure in the cortex of adult and old mice
- In earlier work, we have developed a novel approach to automated segmentation of synapses in Electron Microscopy (EM) image stacks. It relies on image features specifically designed to take spatial context into account. These features are used to train a classifier that can effectively learn cues such as the presence of a nearby post-synaptic region. As a result, our algorithm successfully distinguishes synapses from the numerous other organelles that appear within an EM volume, including those whose local textural properties are relatively similar.In this talk, I will show that this approach has now reached the level of maturity required to obtain biologically relevant results. More specifically, we used it to compare the synapses in layer 1 of the somatosensory cortex in aged and adult mice. Our analysis indicates that the connectivity in ageing brain alters beyond a simple loss of synapse numbers. It also involves changes in size of inputs and the excitatory/inhibitory balance. This indicates a shift towards a more stable connectivity with fewer transient connections and reduced plasticity of the cortical circuits.
- Jeroen van der Laak – Challenges of using computer vision for evaluating whole-slide pathology images
- Although currently feasible, not many pathology laboratories make the transition to a full digital workflow. High costs and regulatory issues hamper wide scale introduction of whole slide image scanners. Availability of validated computer aided diagnosis (CAD) algorithms may dramatically change this situation. These algorithms have the potential to extract clinically meaningful information from scanned tissue sections in a reproducible manner. However attractive, in this talk I will show the difficulties in this new field of research. Large image sizes and a high degree of variability in images are some of the factors making this a challenging field. It remains to be seen whether currently available pattern recognition techniques are sufficiently powerful to address these issues.
- Gonzalo de Polavieja – Tracking animals in groups from video
- We study collective animal behavior. Our starting data is animal tracks from video. We devised a method, that we named idTracker, based on identification of each animal in the group from its image. Tracking by identification eliminates error propagation and it is not necessary to perform manual corrections.. I will talk about idTracker, its extensions to 3D and large groups, and how we use the data for analysis and modelling of group behavior.
- Jens Rittscher – Monitoring Complex Biological Environments
- Building on recent advances in computer vision and machine learning we are now in the position to monitor complex biological environments and events in the same way are analysing natural scenes. While challenges remain, algorithms for cell segmentation and tracking have matured significantly and can now be used in more routine high-throughput settings. Improved microscopy and imaging platforms not only allow us to image subcellular events at high spatial and temporal resolution, we can now image large tissue sections and capture how various different proteins modulate the cellular microenvironment. Enabled by advances in cell culturing technologies 3D cultures can restore specific biochemical and morphological features that are similar to their in vivo counterparts. This holds the potential for improving relevance of in vitro studies, improving our ability to predict what occurs in vivo. We are now working towards establishing the spatial and temporal context for biological events and processes. Quantitative image analysis methods are necessary for monitoring the tissue formation process and enabling longer duration time-lapse imaging. Quantitative imaging can be used very effectively to analyse the cell-to-cell and cell-to-matrix interactions that characterize the microenvironment as well as migration and invasion mechanisms. A more ambitious goal is the analysis of collective cell migration, which plays a crucial role in development and disease progression. The talk will provide examples on how quantitative imaging will advance our understanding of biological mechanisms.
- Thomas Walter – Bioimage Informatics for Phenomics – Exploring morphological phenotypes and spatial organization by large-scale screening approaches
- High-throughput next generation sequencing techniques and methodological advances in the field of bioinformatics have revolutionized human and animal genome researches and have given us unprecedented insights into the organization, control and variation of genomes and transcriptomes. More recently, these efforts have been paralleled by advances in the field of High Content Screening, allowing us to systematically study cellular phenotypes as a consequence of gene silencing or drug exposure by image-based approaches. Importantly, such large-scale imaging approaches are complementary to most molecular approaches in several ways: images typically allow one to analyze spatial arrangements of cells, compartments or molecules, to perform analysis at the single cell level and thereby to study cellular heterogeneity and they are informative about morphological phenotypes, which are otherwise impossible to analyze. Here, I will report on recent advances and new challenges in this field.As it is the case for sequencing data, large-scale image data sets constitute a scientific resource of great but often-underappreciated value to make new discoveries and to trigger methodological advances. Here, I will show how we can make use of a screen initially designed to study cell division [1] in order to analyze cellular motility [2]. For this, my team has developed a methodological framework to quantitatively study cell movement from live cell imaging data, combining cell tracking and cell trajectory analysis by unsupervised clustering. Large-scale HCS data sets can also be used as a reference set for parallel drug screening, where targeted pathways are inferred from the phenotypic comparison between drug exposure and gene silencing experiments. For this, we have developed a distance metric based on the Sinkhorn divergence that allows injecting prior knowledge on phenotypic morphologies and their biological similarity.In addition to survey morphological phenotypes, we can also systematically analyze spatial configurations with large scale screening approaches. Here I will report on a recently developed framework for the analysis and simulation of single molecule Fluorescence In Situ Hybridization (smFISH) data, informative about the spatial distribution of transcripts in cells.[1] Neumann, B., Walter T. et al (2010). Nature, 464, 7289:721-7.[2] Schoenauer Sebag, A. et al. (2015). Bioinformatics, 31, i320–i328.