Algorithmic Information Theoretic Prediction & Discovery
Our approach uses algorithmic information theory to analyze results from accurately segmenting and tracking objects in the image sequences. Multi-target tracking is used to establish temporal correspondences between the segmented objects. A multi-resolution algorithmic information theoretic distance measure computes pairwise comparisons between time courses of object feature values. This results in an ensemble of distance matrices that are further analyzed using a class of semi-supervised spectral learning techniques. Our approach is semi-supervised. If final outcomes are unknown, it functions as a "tool for discovery," and uses a clustering model based on the notion of randomness deficiency from algorithmic statistics to capture as much meaningful information from image sequence data as possible. If final outcomes are known it becomes semi-supervised, incorporating both labeled and unlabeled data for the pair-wise comparisons and then using cross validation together with supervised spectral learning for the most accurate predictions based on behavior.
Link to our our biological collaborators in McGill Canada: [[1]]
Algorithmic Information Theoretic Prediction example
Download
The algorithmic information theoretic prediction and discovery tools have not been integrated into the FARSIGHT build environment. As a temporary measure, source code and Win 32 binaries can be downloaded from Andrew Cohen - UWM
Usage
I. Segment and Track (folder: .\segment and track)
The first step in the process is to segment and track the image sequence data. This step is typically application-specific; however there are some standard software packages that can help [2]. This release includes segmentation and tracking for vertebrate retinal progenitor cells (folder "segment and track") and also for a simulated data set (folder "phantom 2"). The results of the segmentation and tracking form the input to the algorithmic information theoretic prediction and discovery tools. The Matlab file GenerateDistanceMatrices\ExportTS.m illustrates how to export Matlab data into the text format used by the algorithmic information theoretic prediction and discovery tools. All the objects being analyzed must be exported to a single file (except in the case of the real-time solution described below).
Reference
Andrew R. Cohen, Christopher Bjornsson, Sally Temple, Gary Banker, and Badrinath Roysam, “Automatic Summarization of Changes in Biological Image Sequences using Algorithmic Information Theory” (in press, ePub available) IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008/2009.
