Dr. Melanie A Sutton

Snakes have been widely used for object tracking, shape detection, and segmentation of an area of interest within image data. A snakes algorithm uses an energy minimization approach to deform an initial boundary or curve so that it traces along the contour of a shape in an image. However, a major disadvantage of the algorithm is that it requires users to draw the initial boundary in the image of an object, which is not feasible when large number of images need to be processed or when user-introduced bias in the selection of the initial boundary may influence the accurate detection of objects. This paper reports on an algorithm for the automatic detection of a region of interest that utilizes a snakes algorithm for image segmentation. It specifically combines multiple image processing and screening techniques to build a pipeline of processing steps that produces the initial boundary of a region in an image for the initialization of a snakes algorithm. The approach has been evaluated on X-ray images of the striped burrfish to detect the swimbladder as a region of interest. Results from the fully automated algorithm are compared against ground truth values and semi-automated algorithm results.

This paper presents the development of an application for digital image processing of x-rays of stingrays. In order to measure the rate of absorption at varying temperatures, stingrays were fed detectable ball bearings. Using image processing techniques similar to digital mammography before the development of PACS, ball bearings and anchor points are detected in 111 scanned x-ray images of stingrays. Automatic measurements of distance between the ileum and the ball bearings are taken. The application incorporates image processing techniques and a graphical user interface. The development of the application to automate the measurement process is discussed, as well as its potential use, and its modification for use for different species of aquatic life.

This project represents an interdisciplinary approach to integrating computational methods into the knowledge-discovery process associated with understanding biological systems impacted by the loss or destruction of sensitive habitats. We specifically
developed bioinformatics tools for the study of (1) beach mouse communities and (2) marginal fish habitats. Data mining was used in these projects to intelligently query databases and to elucidate broad patterns that facilitate overall data interpretation. Visualization techniques that were developed present mined data in ways where context, perceptual cues, and spatial reasoning skills can be applied to uncover significant trends in behavioral patterns, habitat use, species diversity, and community
composition.

Work presented here focuses on employing wavelets, multi-resolution guided fuzzy c-means (FCM), and coarse-to-fine feature analysis for rapid detection of microcalcification clusters in uncropped images. FCM segmentation is guided through a multi-resolution approach to rapidly distinguish medically relevant tissue from background. Sets of overlapping sub images, containing only relevant tissue, are extracted from the image for further high-resolution analysis. A minimum number of features are used in a simple fuzzy system to
detect candidate microcalcifications. This coarse detection is shown to provide high detection rates while minimizing the data points requiring further feature analysis. Feature extraction and classification is performed with a radial basis function (RBF) neural network. Cluster analysis provides final detection.

We describe the use of segmentation with fuzzy models and classification by the crisp "k-nearest" neighbor (knn) algorithm for assisting breast cancer detection in digital mammograms. Our research utilizes images from the "Digital Database for Screening Mammography" (DDSM). We show that supervised and unsupervised methods of segmentation, such as k-nn and "fuzzy c-means" (FCM), in digital mammograms will have high misclassification rates when only intensity is used as the discriminating feature. Adding window means and standard deviations to the feature suite (visually) improves segmentations produced by the k-nn rule. While our results are encouraging, other methods are needed to detect smaller pathologies such as microcalcifications.

This paper presents a method for image understanding that combines a fuzzy pixel-based feature extractor with a novel, multiple prototype classifier to detect and interpret small targets in LADAR intensity images when very few pixels on target are available. The method is based on the fuzzy c-means clustering algorithm (FCM). Prototypes are appended ta an unprocessed image and low-level attributes of each pixel in the combined image are computed from the 8-neighbor pixels using PCM in 9 dimensions with 5 classes. A feature vector is then extracted from each prototype using a centered n x m window. The class membership vectors of the labeled prototypes are compared to the resulting class membership vectors of each unlabeled pixel to generate a set of confidences of a pixel's membership in the prototype classes. The fuzzy partition produced by PCM retains spatial integrity of each pixel label vector and relates the pixel level information contained in the partition to pixels in the data to be labeled. The method exhibits good behavior for images that do not contain any of the original prototype targets.

This paper describes our work in enhancing and analyzing digital mammograms from the Digital Database for Screening Mammography (DDSM). The DDSM will ultimately contain 3000 cases and provides a unique opportunity for researchers from around the world to compare results on a large, diverse data set. However, the size of the database and images within it require careful consideration of memory limitation issues, display device constraints, etc. We address research problems connected with the modification and application of existing fuzzy modeling approaches to this digital mammography domain. Segmentation and edge detection are used as benchmark applications for the comparisons we make.

We study classifier fusion by the fuzzy template (FT) technique. Given an object to be classified, each classifier from the pool yields a vector with degrees of “support” for the classes, thereby forming a decision profile. A fuzzy template is associated with each class as the averaged decision profile over the training samples from this class. A new object is then labeled with the class whose fuzzy template is closest to the objects’ decision profile. We give a brief overview of the field to place the FT approach in a proper group of classifier combination techniques. Experiments with two data sets (satimage and phoneme) from the ELENA database demonstrate the superior performance of FT over: a version of majority voting; aggregation by fuzzy connectives (minimum, maximum, and product); and (unweighted) average.

The premise of our most recent work is that a.recognition system can and should incorporate both the symbolic labeling of the potential functionality of an object and the steps to confirm said functionality through interaction. Hence, the task at hand is as follows. A researcher selects an object and places it in the observation area of a robot arm. An initial intensity and range image are acquired. This initial state is the input to a two-stage recognition system which first performs the symbolic labeling of the object's potential functionality and produces a plan for interaction for the object. The second stage involves the interaction tests, guided by the plan for interaction, to confirm the object's functional use in a task. This paper describes the current state of the system.