CBA Seminars during Spring 2023

Date	Time	Title	Speaker
2023-01-16	14:15-15:00 Theatrum Visuale	Uncertainty Estimation of Semantic Segmentation of MR-Linac Prostate Cancer Images Prostate cancer radiotherapy treatment makes use of medical deep learning methods in many steps of the workflow. Instead of focusing solely on the performance metrics of the deep learning model attention must also be given to the uncertainty of the model output. We applied the Monte Carlo dropout method for estimating the uncertainty of a semantic segmentation model of MR-Linac prostate cancer images. Concrete dropout was added to the middle layers of a simple U-Net to obtain multiple samples from a stochastic distribution for each input image. Epistemic uncertainty and predictive uncertainty were captured in uncertainty maps by using mutual information and predictive entropy. The predictive entropy maps revealed that the wrongly predicted pixels correspond to high estimated uncertainty values.	Marissa van Lente Eindhoven University of Technology, The Netherlands
2023-01-23	14:15-15:00 Theatrum Visuale	Research in the micro-group I will give an overview of the research that has been going on in the 'my' research group since I joined the Dept. of IT in 2011 as lecturer in Quantitative Microscopy. I will also talk a bit about efforts to find a way forwards, including both joys and disappointments. I will also try to include a bit of ideas for the future, and I'm hoping to give you some insights to my joy for being a researcher in the multi-disciplinary domain of bioimage informatics.	Carolina Wählby
2023-01-30	14:15-15:00 Theatrum Visuale	Uniform Polyhedra – or What’s in the Vitrines on Floor 4? The history of the 75 uniform polyhedra starts with early humans and ends 1975. In this seminar, I will tell that story. Early efforts are attributed to Plato and Archimedes – doubtful – while Johannes Kepler 1600s, Louis Poinsot 1800s, Donald Coxeter et al., 1900s really did contribute to the story. Uniform polyhedra will be defined and sorted into different types. The construction of a number of uniform polyhedra will be discussed and presented. As an extra benefit, you will also understand that most mathematicians do not die young.	Gunilla Borgefors
2023-02-06	14:15-15:00 Theatrum Visuale	Towards an Improved Oral Cancer Classification via Multimodal Image Fusion of Brightfield and Fluorescence Micrographs Diagnosis of oral cancer from brush biopsies combined with automatic detection has the potential to be a non-invasive cost-effective approach for early detection and intervention. In this project, we have investigated the potential for improving the accuracy of the automated classification of cancer in the individual cells by using multimodal imaging and analysis, incorproating both brightfield microscopy (which is commonly used) and also fluorescence microscopy of the Papanicolaou stained samples. We suggest an approach that involves the multimodal image acquisition, cell nuclei detection, cross-modal registration of cell nuclei to find matching image patches, and a multimodal deep learning classifier with fusion. We observe a substantial improvement in performance from incorporating the additional modality on a dataset of 8 cancer patients and 11 healthy people.	Johan Öfverstedt
2023-02-13	14:15-15:00 Theatrum Visuale	Sharing synthetic medical images Deep learning can solve very complex tasks in image processing, but requires large annotated datasets. In medical imaging it is difficult to share images due to ethical regulations and GDPR. During the last years, GANs and diffusion models have been trained to create very realistic synthetic images. In this presentation I will talk about generating synthetic medical images, about training networks with synthetic images, and what the legal implications are for synthetic images.	Anders Eklund Dept. of Biomedical Engineering, Linköping University Host: Andreas Hellander
2023-02-20	14:15-15:15 Theatrum Visuale	Docentship lecture Ultrasound Imaging: Fundamentals and Applications Medical image analysis aims to generate value from medical images, and a fundamental understanding of the acquisition of such images is essential for devising processing methods and hypotheses utilized therein. Among such modalities, ultrasound imaging is a unique option being safe, real-time, portable, and affordable; and ultrasound is used in many clinical applications, including diagnostic decision making, pre-operative planning, surgical guidance, monitoring, and screening for diseases. This lecture will introduce the basics of ultrasound imaging and will present different ultrasound imaging techniques and contrast mechanisms; from conventional approaches to recent innovative methods such as in the fight against cancer, detecting aging muscles, and functional brain scans.	Orcun Göksel
2023-02-21	14:15-15:00 Theatrum Visuale	C-grain company presentation C-grain is an Uppsala based company developing and providing instruments for grain analysis based on imaging and image analysis including AI. See this video for an introduction	Moa Källgren, Development Manager Customer Projects; Jonas Persson, Software Development Manager; Mikael Åstrand, Managing Director
2023-02-27	14:15-15:00 Theatrum Visuale	Stability of image-reconstruction algorithms In many imaging modalities, images are only obtained by applying machine-learning algorithms to indirect measurements (e.g., irregular measurements of the Fourier domain in MRI, or Poisson events related to the sinogram in PET). Image-reconstruction algorithms select the best reconstruction among an infinite collection of alternatives by using model-based or data-driven prior information. Clearly, the reliability of these algorithms is crucial to ensure their adoption in regular practice. Despite this, the stability of many modern methods has not been studied quantitatively, and state-of-the-art neural-network-based algorithms exhibit serious practical stability problems. In this talk, I will summarise the state of the field and present our new results on the stability of lp-regularized linear inverse problems for p in (1, infinity). More details can be found in our very recent publication.	Pol Del Aguila Pla Biomedical Imaging Group, EPFL Lausanne Host: Nataša Sladoje
2023-03-06	14:15-15:00 Theatrum Visuale	A simple and fast template matching by nearest-neighbour based feature distribution matching Template matching is a fundamental problem in computer vision and has applications in various fields like object detection, object tracking, image registration etc. The current state-of-the-art methods rely on nearest-neighbour (NN) matching of a small subset within the template to the query image patches. The methods have been shown to perform better in the cases of occlusions, changes in appearance, illumination variations, and non-rigid transformations. However, NN-based algorithms rely on computing the NNs for each pixel in the template which becomes time-consuming at higher image resolution and also prohibits the use of high-dimensional feature representation as the computational cost grows in both cases. In this work, we present an NN-based distribution matching method which relies only on a few representative features in the template. We also introduce filtering in the nearest neighbour field (NNF) which can be used to consider the deformations implied by the NNF. We show that state-of-the-art performance can be achieved in low-resolution data and our method outperforms previous methods at higher resolution.	Ankit Gupta
2023-03-13		No seminar SSBA/SSDL Symposium
2023-03-20	14:15-15:00 Theatrum Visuale	CANCELLED
2023-03-27	14:15-15:00	Using Machine Learning to Discover and Characterise Atmospheres of Exoplanets Investigation of planets orbiting stars other than the Sun (exoplanets) is one of the most exciting and active research directions in modern astronomy. The current frontier in this research area is to advance from a mere detection of exoplanets to detailed physical characterisation of their atmospheres and surfaces. A fundamental observational basis of this research is detection and interpretation of spectra of exoplanet atmospheres. In this talk I will discuss our ongoing research aimed at detecting and separating exoplanet spectra from the far stronger radiation of host stars using high-resolution transmission spectroscopy. I show that this task can be expressed as the problem of finding regular patterns in noisy images, to which various machine learning approaches can be applied. I will present preliminary results of applying one of such approaches to simulated exoplanet observations.	Oleg Kochukhov, Dept. of Physics and Astronomy Host: Nataša Sladoje
2023-04-03	14:15-15:00	Learning-based prediction, representation, and multimodal registration for bioimage processing Microscopy and imaging are essential to understanding and exploring biology. Modern staining and imaging techniques generate large amounts of data resulting in the need for automated analysis approaches. Many earlier approaches relied on handcrafted feature extractors, while today's deep-learning-based methods open up new ways to analyze data automatically. Deep learning has become popular in bioimage processing as it can extract high-level features describing image content. The work in the thesis explores various aspects and limitations of machine learning and deep learning with applications in biology. Learning-based methods have generalization issues on out-of-distribution data points, and methods such as uncertainty estimation and visual quality control can provide ways to mitigate those issues. Furthermore, deep learning methods often require large amounts of data during training. Here the focus is on optimizing deep learning methods to meet current computational capabilities and handle the increasing volume and size of data. Model uncertainty and data augmentation techniques are also explored. PhD rehearsal	Nicolas Pielawski
2023-04-10		No seminar Easter Monday
2023-04-13	15:15-16:00, room 106157	Predictive Models for Computational Pathology and the analysis of RNA localization patterns In the field of bioimaging, we dispose of the technological tools to perform imaging experiments at an unprecedented scale. For instance, High Content Screening (HCS) allows us to systematically study the spatial distribution of proteins and RNAs inside cells, involving hundreds of thousands of experiments in a single project. The study of stained tissue slides allows us to study morphological patterns and tissue changes in response to disease. Both HCS and digital pathology thus generate large and complex image data sets informative about cellular and tissular phenotypes and the multi-scale properties of living systems. The method of choice to analyze these challenging datasets is Computer Vision, and Deep Learning in particular. In my seminar, I will present our recent developments for the analysis of subcellular RNA localization patterns, involving point cloud analysis, and for the analysis of stained tissue slides in view of predicting disease relevant variables, and involving self-supervised and Multiple Instance Learning. (Extra seminar in connection to the PhD thesis defense by N. Pielawski)	Prof. Thomas Walter, Ecole des Mines de Paris Host: Carolina Wählby
2023-04-17	14:15-15:00	Insight to the BioImage Informatics Facility (BIIF) The BioImage Informatics Facility provides support and training to perform state-of-the-art analyses on image data of researchers all over Sweden. Our experts can help deploy computational methods using computer vision, machine learning, and bioinformatics to analyze images. In the seminar we will give more insight into the activities of BIIF. We will present one user project in more detail: the aim here is to analyze CT-images of fungus-infected wood logs. We will also present the newest development in the TissUUmaps project.	Anna Klemm, Christophe Avenel, Fredrik Nysjö, Jonas Windhager
2023-04-24	14:15-15:00	Image Processing and Analysis Methods for Biomedical Applications (PhD rehearsal) With new developments and rapid acquisition times medical images can nowadays be acquired at a larger scale than ever before. With increased amounts of available data, computerized image processing and analysis methods play an important supporting role in diagnostic pipelines, enabling swifter examinations and decision-making. In this thesis we develop and improve methods for deep learning-based image segmentation under weak supervision and patch-based training, instance-level cross-modality image retrieval and large-scale statistical analyses, with potential applications in imaging-based diagnostic pipelines.	Eva Breznik
2023-05-01		No seminar May Day
2023-05-08	14:15-15:00	Segmentation of in situ transcriptomics data In recent years, spatially resolved multiplexed in situ transcriptomics (IST) techniques have undergone significant advancements. These approaches enable the mapping of mRNA molecules directly within tissue samples, which allows researchers to dissect and analyze cell-type heterogeneity in a spatial context. Essentially, the data generated from these techniques can be described as a set of markers, with each marker representing a different mRNA molecule. Each marker has two attributes: a spatial location indicating the mRNA molecule's position within the tissue and a categorical label indicating its type. To understand the biological mechanisms at play, grouping these markers into components like cells and cell types is useful. The process of grouping markers into cells and cell types can be likened to traditional image analysis concepts such as instance and semantic segmentation. In this seminar, I will provide an overview of various methods for grouping mRNA markers into individual cells and cell types. Additionally, I will introduce two simple techniques developed by myself and Andrea Behanova for these purposes.	Axel Andersson
2023-05-15		No seminar EMBL-SciLifeLab Data Science Workshop
2023-05-22	14:15-15:00	Multimodal Digital Cytology with CytoBrowser - your collaborative Whole Slide Image / Virtual microscope environment I will present some new features of the CytoBrowser software for fast and accessible collaborative online visualization, assessment, and annotation of very large (whole slide) microscopy images, including support for multiple modality layers and machine learning prediction import. I will also show some examples on how we are aiming to improve Deep learning based Oral Cancer detection by combining brightfield and fluorescence microscopy.	Joakim Lindblad
2023-05-29	14:15-15:00	Representation Learning and Information Fusion - Applications in Biomedical Image Processing In recent years Machine Learning and in particular Deep Learning have excelled in object recognition and classification tasks in computer vision. As these methods extract features from the data itself by learning features that are relevant for a particular task, a key aspect of this remarkable success is the amount of data on which these methods train. Biomedical applications face the problem that the amount of training data is limited. In particular, labels and annotations are usually scarce and expensive to obtain as they require biological or medical expertise. One way to overcome this issue is to use additional knowledge about the data at hand. This guidance can come from expert knowledge, which puts focus on specific, relevant characteristics in the images, or geometric priors which can be used to exploit the spatial relationships in the images. This thesis presents machine learning methods for visual data that exploit such additional information and build upon classic image processing techniques, to combine the strengths of both model- and learning-based approaches. The thesis comprises five papers with applications in digital pathology. Two of them study the use and fusion of texture features within convolutional neural networks for image classification tasks. The other three papers study rotational equivariant representation learning, and show that learned, shared representations of multimodal images can be used for multimodal image registration and cross-modality image retrieval. PhD rehearsal	Elisabeth Wetzer
2023-06-05		No seminar Bridge Day for June 6
2023-06-12		No seminar PhD thesis defense Elisabeth Wetzer
2023-06-13 (Tuesday!)	13:15-14:00 Theatrum Visuale Zoom-link here	AI-driven computational pathology Digital pathology is rapidly transforming the workflow in routine diagnostics, enabling the use of computational methods for interpreting the data. Our aim is to enable faster, less subjective and in some cases even more accurate diagnostics through deep learning based computational pathology. Besides enabling decision support for tasks currently done by human experts, computational pathology has the potential for novel discoveries from histopathology beyond the limits of human vision. Our research focus is on data-intensive research questions in cancer research and computational pathology. The presentation will cover some of our recent work in AI-driven cancer diagnostics, reconstruction and visualization of histological data in 3D, and in developing virtual staining for unstained tissue.	Pekka Ruusuvuori, Bioimage informatics group leader, Inst. of Biomedicine, University of Turku Host: Nataša Sladoje
2023-06-13 (Tuesday!)	14:15-15:00 Theatrum Visuale Zoom-link here	Dimension reduction: in general, and in single cell data analysis Visualizing high-dimensional data is an essential step in exploratory data analysis, and it has become routine to show UMAP embeddings, including of single cell data. But how canonical is the use of UMAP, and what do we know about its inner workings? The first part of the talk will provide a brief tutorial overview over strategies for dimension reduction, including neighbor embeddings such as UMAP and tSNE, force-directed layout and auto encoders. In a second part, I will highlight our own contributions, including work on the precise relation between t-SNE and UMAP [Neurips 2022, ICLR 2023], tree-biased auto-encoders [Bioinformatics 2021] and geometric auto-encoders [ICML 2023]. Finally, using the example of STAGATE [Dong&Zhang 2022] I will discuss how spatial cues can inform dimension reduction in spatial transcriptomics. Altogether, and especially in spatial omics, the problem of dimension reduction is anything but "solved".	Fred Hamprecht, Professor in Image Analysis and Learning, Heidelberg University. Host: Nataša Sladoje
2023-06-19	14:15-15:00 Theatrum Visuale	Deep learning-based species classification of time-lapse of bacteria growing in microfluidic chip chambers by phase-contrast microscopy (Half-time seminar) Reliable detection and classification of bacteria and other pathogens in the human body, animals, food, and water is crucial for improving and safeguarding public health. For example, the species and its antibiotic susceptibility are essential to treating a bacterial infection. In this half-time seminar I will summarize what I have been working on in my PhD project, namely investigating deep learning-based computerized image processing methods to classify the species of bacteria growing in microfluidic chip chambers using only phase-contrast microscopy, using both still images and time-lapse "video" data. Zoom Link: https://uu-se.zoom.us/j/62644248488 Reviewer: Orcun Göksel	Erik Hallström