Biomedical Image Analysis
Tuesdays, 14:00 - 15:30
DBE - University of Basel
|20. Sep 2016||Visit Rectorate, University of Basel/Switzerland||Ph. Cattin/G. Rauter/G. Jost|
|27. Sep 2016|
Computational approaches to diseases of the central neural system: In this talk I will summarize efforts and experience conducted to develop decision-making support systems aiming at better quantifying diseases of the central neural system. Special emphasis will be given to the field of neuro-oncology as well as neuro-vascular diseases such as ischemic stroke. A second part of the talk will provide insights into different applications, such as radiomics and response assessment in neuro-oncology.
|Prof. Dr. Mauricio Reyes, University of Berne|
|4. Oct 2016||Fast Convolutional Neural Networks for Graph-Structured Data: Convolutional neural networks (CNNs) have greatly improved state-of-the-art performances in computer vision and speech analysis tasks, due to its high ability to extract multiple levels of representations of data. In this work, we are interested in generalizing convolutional neural networks from low-dimensional regular grids, where image, video and speech are represented, to high-dimensional irregular domains, such as social networks, telecommunication networks, or words' embedding. We present a formulation of CNNs in the context of spectral graph theory, which provides the necessary mathematical background to design localized convolutional filters on graphs. More importantly, the proposed technique offers the same computational complexity than standard CNNs, while being universal to any graph structure. Numerical experiments on MNIST and 20NEWS demonstrate the ability of this novel deep learning system.||Dr. X. Bresson, EPF Lausanne|
|11. Oct 2016||Canceled|
|18. Oct 2016||(MICCAI in Athens)|
|25. Oct 2016|
|1. Nov 2016, Shifted to Mon. Nov. 28th 2016, 12h30-14h|
Image-based modelling of tumor growth: Glioma is the most frequent primary brain tumor and extensive neuroimaging protocols are used to evaluate the progression of the disease and the success of a chosen treatment strategy. This gives rise to large and complex multimodal data sets, and extracting diagnostic information across different clinical imaging modalities and along time poses a significant problem when analysing these data. We provide an overview of the state of the art in brain tumor image segmentation, and present a generative model of tumor growth and image observation that describes the tumor evolution at the macroscopic imaging level. Model personalization relies on a forward model of the patho-physiological process adapted to organ geometries together with image likelihood functions, and an efficient Bayesian inference approach. We illustrate the application of the tumor growth model in radiation therapy.
|Prof. Dr. Bjoern Menze, TU Munich|
|8. Nov 2016||Quantitative Medical Image Analysis for Personalized Decisions: The advances in acquisition technologies and the increasing use of patient-specific information in clinical practice are giving computational image analysis and modelling important roles in healthcare. This presentation will be about my overall research on medical image computing. First, I will present different components of image analysis with examples from my own research within this big picture. Then, I will present in detail three different topics: semantic image parsing, group analysis and predictive modelling for diagnosis and risk assessment. Then I will conclude with open problems and future research directions.||Prof. Dr. Ender Konukoglu, ETH Zurich|
|15. Nov 2016|
Data Compression with DRain: Current imaging techniques allow the examination of physical objects at unprecedented resolutions, leading to very large datasets. DRain is a software technology that enables applications to inspect and visualize a 100+ GB dataset at interactive rates, running on supercomputers as well as on thin clients and notebooks. DRain leverages the diverging gap between the I/O bandwidth and the peak performance of contemporary CPUs. A single-byte I/O transfer can hide in the order of thousands of floating point operations. This “free budget of FLOPs” is employed to decorrelate data by performing a 3D wavelet transform. Special second-generation wavelets and a wavelet-aware codec have been designed to achieve very competitive compression rates while accurately preserving the signal at coarser resolutions. After the lossless data compression, DRain is able to supply dataset previews to the applications at unrivaled performance-accuracy points. The generated previews are of incremental fidelity and eventually reproduce the exact dataset.
Dr. Diego Rossinelli, R&D at Lucid Concepts AG
|22. Nov 2016||Sparse Bayesian Image Registration: Among the limitations of current registration algorithms are the issue of tuning the registration parameters and that of estimating the uncertainty of the transformation. In this lecture, we introduce a sparse Bayesian non-rigid registration method which tries to address both problems. Image registration is posed as a linear regression problem with 2 sets of priors. The former allows a sparse selection of displacement bases defined as Gaussian distributions while the latter enforces the smoothness of the resulting displacement field. The proposed approach extends the Relevance Vector Machine formulation in many ways and inference of parameters is performed through a Variational Bayes (VB) approach. The sparsity constraint allows to estimate the posterior covariance matrices and therefore the uncertainty of the transformation. Finally this estimation of uncertainty based on this VB inference has been compared with an estimation of the true posterior distribution through stochastic (MCMC) sampling.||Prof. Dr. Hervé Delingette, INRIA/France|
|29. Nov 2016|
Optimal segmentation models for fluorescence microscopy images: Detecting and segmenting objects in fluorescence microscopy images becomes more accurate and robust by including prior knowledge about the labeled objects and about how the images have been acquired. This is the basis of model-based image analysis. But how accurately is one possibly able to segment an object, given the information available in an image? Which algorithms extract the maximum amount of information from an image? This is the topic of optimal image analysis. We show information-theoretic bounds for segmentation of extended objects and filaments, and propose model-based algorithms that asymptotically reach these bounds. This enables both user-friendly and optimal segmentation, as implemented in the open-source software “MOSAICsuite” for Fiji and ImageJ.
|Prof. Dr. I. Sbalzarini|
|6. Dec 2016||Recent works on Compressed Ultrasound beamforming: Classical ultrasound (US) image reconstruction mainly relies on the well-known Delay-And-Sum (DAS) beamforming for its simplicity and real-time capability. However, DAS requires an extensive number of samples and delay calculations to obtain high-quality images. Compressed ultrasound beamforming (CUB) proposes an alternative to DAS based on the compressed-sensing (CS) framework which aims at reducing the data rate. CS demonstrates that a signal can be perfectly recovered from fewer samples than required by the Nyquist rate if some properties of both the signals under interest and the acquisition system are respected. In order to account for these properties, CUB redesigns both the acquisition and the reconstruction of US images and leads to high quality reconstruction with less than 20% of the data required by DAS. In the talk, some basic principles of CS and US will be introduced. Then, we will describe CUB in light of the CS framework introduced before. Finally, benefits of CUB will be demonstrated through simulation and in vivo experiments.||Prof. Dr. Jean-Philippe Thiran, EPF Lausanne|
|13. Dec 2016|
Extracting structure from large-scale neural recordings: As the size and complexity of neural datasets continue to grow, there is an increasing need for scalable approaches to extract knowledge from these data. In this talk, I will discuss my recent work in developing methods to uncover the structure of neural systems from both X-ray microtomography and serial two photon tomography datasets. My aim is to provide an introduction to a number of modern neural datasets and discuss some of the challenges that we now face in developing learning algorithms for these data.
|Dr. Eva Dyer von der Northwestern University, Evanston, US|
|20. Dec 2016||Exam||Prof. Dr. Philippe Cattin|