Multidimensional Image and Data Analysis
Multispectral Microscopy
Project Members: James Thigpen, Jason Wu
Spectral imaging combines spectroscopy and imaging to acquire a spectral image that measures the complete spectrum of a sample at each point in the imaging plane. Spectral images are three-dimensional cubes of data, comprised of a series of 2D images, one for each of the wavelengths sampled. We have prototyped a spectral microscopy system capable of imaging an entire slide surface to measure chemical characteristics of a specimen, thereby adding capabilities beyond traditional morphological analysis.
We are developing methods for spectral cytology of fine needle aspirated cells leading to differential analysis of malignant and benign tumors without resorting to histological assessments. Specific methods for cell segmentation, feature extraction, spectral unmixing, and visualization are being developed.
|
|
| Automated Spectral Imaging System | Differential Absorption in Aspirated Cells |
Multimodality Biometric
Project Members: Joseph Mathew
We are developing a multimodal biometric system based on a novel fusion strategy that would be capable of combining multiple signatures to alleviate the limitations of unimodal systems. Our approach uses a two-stage probabilistic integration that combines multiple features from distinct biometric traits and the respective matching obtained by each biometric to evaluate a final match. The overall strategy of the proposed framework is represented in the figure below and includes the following stages: a) multiple feature sets are computed from each biometric modality and modeled to find a relationship or mapping between the feature set and the class label (the enrolled individual). Each mapping provides an initial estimate of classification probability of the observed features corresponding to the class label at the expense of a high false alarm rate. This allows us to explicitly model the problems associated with noisy data and biometric drifts; b) focused analysis of the derived mapping is performed in the following stage. The objective at this stage is to identify the degree of participation of each feature towards a reliable and robust class label association. This is done by estimating the confidence measure of each mapping in successful classification of the observed biometric under varying operational conditions; and c) the final stage is a decision integration from the lower levels to identify the combined decision for the modeled class label from all possible labels presented by the individual classifiers.
Perceptual Visualization
Project Members: Vyom Munshi
Visualization is critical in our ability to select data transformations necessary for common data mining tasks. 2D scatter plots and histograms have been used extensively for this process, but are limiting due to traditional display and visualization systems which restrict the number of dimensions that can be simultaneously visualized. They fail to capitalize upon the humanís ability to discern shades of gray, color, size, shape, and texture, of objects.
To address these issues, we are developing high-dimensional visualization tools integrated with data projection techniques that projects high-dimensional data to perceptual attributes such as size, shape, color, texture, position, orientation, brightness, etc.