Program

9h45 Welcome coffee and Opening

10h15 Paul Eilers, Emeritus Professor in Statistics, Erasmus University Medical Center, The Netherlands.

The Power of Penalties. A penalty is an extension of a statistical model that forces its solution in a desired direction. A penalty can be combined with all types of objective functions for the fit of the model. The sum of squares is familiar, but also (log-)likelihoods for counts and binary data can be used, or robust and asymmetric functions.  Familiar penalties constrain the size of parameters (as in ridge regression) or their roughness (as in smoothing). For many tasks in image analysis penalties are extremely effective. I will describe a number of applications, like noise removal, baseline and range correction and sparse deconvolution.

11h15 Alberto Ferrer, Professor at the Dept of Applied Statistics, University Politecnica Valencia, Spain

Functional magnetic resonance image analysis using multivariate curve resolution methods: developing cancer biomarker. A quantitative imaging biomarker (QIB) can be defined as an extracted feature of medical imaging, which can be measured objectively (ie, give quantitative information) and acts as an indicator of a normal biological process, a disease or a response to a therapeutic intervention. Perfusion and diffusion magnetic resonance (MR) sequential images are used in order to compute these QIB at each pixel location. The talk illustrates how to use chemometrics tools such as multivariate curve resolution (MCR) methods to develop new QIB for cancer diagnosis and compare its performance with classical first principles-based models.

12h30-13h45 Lunch on site

14h00 Harald Martens, External professor, Dept. Engineering Cybernetics, Norwegian U. of Sci. and Technol. (NTNU), Trondheim Norway and Research leader at Idletechs AS,

Hyperspectral Video: Compression and Comprehension. Multichannel spatiotemporal measurements from e.g. hyperspectral imaging and - video in the visible and NIR wavelength ranges require data compression, mathematical modelling, statistical validation, classification and prediction as well as graphical display for human interpretation. This requires insight from many fields of science to be combined. A generic data modelling approach for rational processing of a massive stream of imaging data will be presented. This On-The-Fly combination of signal compression, mathematical modelling, statistical validation and computer graphics renders a high-dimensional data stream interpretable in the compressed form.

15h00 Nicolas Dobigeon, Ass. Prof. in Signal and Communication, Institut National Polytechnique de Toulouse, France

Spectral mixture analysis - Beyond the linear mixing model. Spectral mixture analysis, aka spectral unmixing, is a crucial step while analyzing data provided by various imaging modalities, including astronomy, remote sensing and microscopy. It consists of decomposing the measurements into a set of elementary spectra and quantifying their respective proportions in the observed mixtures. A large majority of the unmixing approaches implicitly or explicitly assume that these measurements result from a linear combination of the elementary spectra. Recent advances, mainly conducted in the remote sensing and geoscience community, have attempted to overcome the intrinsic limitations of this linear mixing model. This talk will propose an overview of these concurrent models, allowing nonlinearities, spectral variability and other mismodeling effects to be taken into account during the unmixing process.

16h00 Closing and Drinks