All CBI seminars take place at 12pm in the CBI Conference Room, Hamerschlag Hall C119. Contact CBI seminar host Gustavo Rohde for more information.
2009 Spring
May 7, 2009, 11am (note time change)
Ann B. Lee
Department of Statistics, Carnegie Mellon University
Diffusion maps with an application to texture discrimination
For naturally occurring data, the dimension of the given input space is often very large while the data themselves have a low intrinsic dimensionality. Diffusion map is a non-linear spectral method for transforming data into a coordinate system that efficiently reveal the geometric structure -- in particular, the "connectivity" -- of the data. In this talk, I will first review the basic ideas behind the diffusion framework and then discuss my current work on texture discrimination by a novel geometry-based metric on distributions. (Part of this work is joint with C. Schafer
February 26, 2009
Zoltan N. Oltvai
Department of Pathology, University of Pittsburgh
Digital Pathology and computational disease modeling
Advances in genomics, imaging, and systems biology promises to contribute significantly to the diagnostic process in pathology. In this talk I will review two of our recent studies that aim to contribute toward this goal.
January 29, 2009
Justin Newberg
PhD student, Dept. of BME, Carnegie Mellon University
Automated Analysis of Subcellular Protein Patterns in Human Tissues
Systematic information on the subcellular distributions of proteins is required for more accurate cell models that can be applied to clinically relevant cases; moreover, such information plays an increasingly important role in medical diagnoses. Given the number of proteins, conditions, and cell and tissue types for which information is needed, there is a critical need for automated, high throughput acquisition and analysis of subcellular location patterns. Automated pattern recognition methods have been shown to be effective at determining protein patterns in limited cell culture datasets. We have adapted these machine learning methods to analyze protein patterns in tissue images. For this purpose, we have used the extensive collection of images in the Human Protein Atlas, which contains over 6000 proteins in immunohistochemically stained tissues. Our initial work on a subset of the Atlas showed that we can determine protein locations across 45 different tissue types with a high degree of accuracy. In this talk, I will discuss various methods- such as classification, clustering, and segmentation- for scaling automated analysis to a larger set of proteins in the Atlas, and I will show preliminary results obtained using these methods.
2008 Fall
December 4, 2008
Charles Jackson
PhD student, Dept. of BME, Carnegie Mellon University
Intelligent Acquisition and Model Building of Fluorescence Microscope Time Series
Fluorescence microscopy is a powerful tool for live cell imaging. However, it suffers from limited spatial and temporal resolution, and photobleaching and phototoxicity limit the duration of acquisition. Furthermore, with an explosion in the amount of data being acquired, visual inspection becomes impractical. In this talk, I will discuss an active learning framework that uses an intelligent acquisition system to efficiently obtain information from a time series. The intelligent acquisition system chooses which pixels to acquire, at which time points to acquire, and when to stop acquisition. I will show some preliminary results obtained from applying this framework to the 3T3 dataset provided by MurphyLab.
November 13, 2008
Jelena Kovacevic
Professor, Depts. of BME and ECE, Carnegie Mellon University
Director, CBI
Active Mask Segmentation of Fluorescence Microscope Data Sets
I will talk about the new active mask (AM) framework and an algorithm for segmentation of digital images, particularly those of punctate patterns from fluorescence microscopy. The AM segmentation framework is suited for digital images. It is based on a local majority voting-based scheme, and can incorporate different forms of the voting function as well as several different functions to skew the voting to obtain a meaningful segmentation result. This framework has multiresolution and multiscale techniques built into it and can be instantiated to segment data of any dimension. We demonstrate the efficacy of the AM through an algorithm for segmenting punctate patterns of cells in fluorescence microscope images. While the theory opens up interesting vistas for research and development, the results demonstrate AM's utility in practice.
October 2, 2008
Chakra Chennubhotla
Dept. of Computational Biology, School of Medicine, University of Pittsburgh
Spectral Methods for Multi-Scale Feature Extraction and Data Clustering
We present a new framework for feature extraction and dimensionality reduction, called Sparse Principal Component Analysis (S-PCA). In this algorithm, we introduce a sparsity constraint on the elements of an orthonormal bases matrix. We show how this constraint can help recover object-specific structure in a low-dimensional subspace in a local, scale-dependent form. The learning algorithm of S-PCA is very simple, consisting of successive planar rotations of pairs of basis vectors. The principal advantages of S-PCA over a standard PCA-based representation include an intuitive understanding of the features underlying the high-dimensional data ensemble and efficiency in computations resulting from a sparse basis representation. We will explore both these themes in the talk. Additionally, using S-PCA we present a new approach to the problem of contrast-invariant pattern detection. The novel contribution of this work is the design of a perceptual distortion measure for image similarity, i.e., comparing the appearance of an object to its reconstruction from the principal subspace.
The other issue that is fundamental to the analysis of naturally occurring datasets is how to cluster items in a dataset using pairwise similarities between the elements. To this end we present a spectral method called EigenCuts. Using a Markov chain perspective, we characterize the spectral properties of the matrix of transition probabilities, from which we derive eigenflows along with their half-lives. An eigenflow describes the flow of probability mass due to the Markov chain, and it is characterized by its eigenvalue, or equivalently, by the half-life of its decay as the Markov chain is iterated. A ideal stable cluster is one with zero eigenflow and infinite half-life. The key insight in this work is that bottlenecks between weakly coupled clusters can be identified by computing the sensitivity of the eigenflow's half-life to variations in the edge weights. The EigenCuts algorithm performs clustering by removing these identified bottlenecks in an iterative fashion. As an efficient step in this process we also propose a specialized hierarchical eigensolver suitable for large stochastic matrices. |