44. EXPLORING MS IMAGING DATA IN A SEMI-SUPERVISED AND INTERACTIVE MANNER

Department: Computer Science & Engineering
Research Institute Affiliation: Graduate Program in Bioinformatics
Faculty Advisor(s): Vineet Bafna

Primary Student
Name: Jocelyne Bruand
Email: jbruand@ucsd.edu
Phone: 858-822-5004
Grad Year: 2012

Abstract
Mass Spectrometric Imaging (MSI), is a molecular imaging technique which allows the generation of topographic 2D maps for a large complement of molecules in the cell. Most bioinformatics approaches have focused on using MALDI MSI as a tool for the discovery of signature markers. One approach is to computationally segment the tissue section into regions, each characterized by a spectral signature. Current methods are black-box tools with little or no user interaction and thus cannot be used in an exploratory way. They also require prior data reduction. We describe an interactive algorithm capable of generating image segments. Previously, we developed a tool which returns the spectral signature for a region of interest (ROI). We expand this tool. By integrating over the m/z values, we determine the likelihood ratio for each MALDI spot belonging to the same cluster as the original ROI. When given several ROIs, the corresponding log-odds images are combined into segmentation maps. We also provide a measure of quality for the segments, allowing the algorithm to run iteratively until quality is satisfactory. The algorithm can also be run on random seed ROIs, providing a completely unsupervised mode. We use our algorithm to create segmentation maps for two MSI datasets: a whole leech embryo dataset and a rat brain dataset. We selected a several partial seed regions for each of the datasets. Our algorithm successfully recovers the molecular signature of each region, as well as finds all other regions with similar signatures. For example, in the leech central nervous system, selecting an anterior ganglion returns the entire CNS, while selecting a posterior ganglion returns a stronger signal in a few of the posterior ganglia. The molecular signatures show that for the m/z values expressed in the CNS, some m/z value show high intensity in the posterior ganglia while others show an intensity distribution throughout the CNS. This is interesting as the leech CNS develops from the head to the tail, and thus anterior ganglia are a few days ahead in development than the tail ganglia. When selecting several region of the cerebral cortex of the rat brain, we find that these return similar spectral signatures and assign high log-odds scores to most of the cerebral cortex. Initializing random seed ROIs also provided good segmentation maps. While our tool can be fully automatic, the user can input information throughout the process to guide tool, making it fully interactive and exploratory.

Related Files:

  1. achemso_TOC.jpg

« Back to Posters or Search Results