摘要:We propose a new method for soft spatial segmentation of matrix assisted laser desorption/ionization imaging mass spectrometry (MALDI-IMS) data which is based on probabilistic clustering with subsequent smoothing. Clustering of spectra is done with the Latent Dirichlet Allocation (LDA) model. Then, clustering results are smoothed with a Markov random field (MRF) resulting in a soft probabilistic segmentation map. We show several extensions of the basic MRF model specifically tuned for MALDI-IMS data segmentation. We describe a highly parallel implementation of the smoothing algorithm based on GraphLab framework and show experimental results.
关键词:MALDI imaging mass spectrometry; hyperspectral image segmentation; probabilistic graphical models; latent Dirichlet allocation; Markov random field