INB-Lunch-Seminar

Mathematical methods for processing hyperspectral imaging data with applications in imaging mass spectrometry

Imaging mass spectrometry is a biochemical measurement technology which allows to establish the chemical composition of a thin sample, e.g. a tissue slice. IMS has several advantages which make this innovative technology a unique and successful tool in cancer biomarker detection and drugs discovery.

An IMS data set represents a hyperspectral image, where for each pixel a high-dimensional mass spectrum is measured. Each mass spectrum dimension represents the abundance of molecules
with this molecular mass. The data set has relatively small number of pixels (e.g. 200x200) but the number of mass dimensions exceeds 10⁴, what requires efficient processing methods.

We consider several computational problems in IMS data processing, namely, (1) representation of a data set with prototype spectra and images, (2) detection of structure in low-resolutional images corresponding to specific masses, (3) unsupervised spatial segmentation with the aim of providing
regions of interest automatically, (4) classification of region-vs-region or sample-vs-sample as well as the use of classification in large-scale studies.

The interesting mathematical methods considered in this talk are: non-negative matrix factorization with graph regularization, spatially-aware clustering, edge-preserving spatially-adaptive image denoising.