2006 Scientific Session Abstracts
E-Poster Presentations
Submission Compliance and Guidelines
Copyright Transfer Form (PDF)
Abstract Archive
A General Framework for Segmenting Imaged Pathology Specimens Using Level-set and Gaussian Hidden Markov Random Fields
Lin Yang, MS1,2 (linyang@eden.rutgers.edu); David J. Foran, PhD2,1; 1Rutgers University, New Jersey, Piscataway; 2Center for Biomedical Imaging & Informatics, UMDNJ-New Jersey, Robert Wood Johnson Medical School, Piscataway.
Context: High-throughput image analysis is an active area of research which holds promise for several key areas of investigative research in pathology and oncology. One of the first, crucial steps in processing is delineation of objects of interest within a given image from background and non-salient regions. All subsequent, higher order abstractions related to the image are reliant upon the speed and accuracy with which the segmentation operation is carried out. While there currently exist a multitude of segmentation strategies, most lack sufficient sensitivity for discriminating among the subtle differences in visual classes contained within a typical pathology specimen. We present a general framework which has proved useful for segmenting cancer tissue microarrays. This new approach was developed after failing to obtain reliable results using standard segmentation strategies. During the course of our experiments it was shown that even Gaussian Mixture models and cluster-based methods alone were unable to provide satisfactory results. In this paper, we describe a level-set based deformable model which exploits the advantages of the Gaussian Hidden Markov Random Field (GHMRF) to produce reliable segmentation results. This segmentation method was utilized to perform both multi-cell and tissue-level analysis.
Technology: The test system consisted of an Intel-based workstation interfaced with a high-resolution Olympus DP70 camera equipped with 12-bit color depth on each color channel, 1.45-million-pixel effective resolution, and a single 
-inch charge-coupled device digital camera (Olympus America, Inc, Melville, NY); an Olympus AX70 microscope (Olympus America, Inc) equipped with a Prior 6-way robotic stage, motorized objective turret (Prior Scientific, Inc, Rockland, Mass); and a magnification changer. The algorithm is currently being modified to support the whole slide scanner.
Design: Image segmentation can be classified as region-based and contour-based approaches. The region-based segmentation approach has a very long history and performs very well in many applications. This class of algorithms includes the mean-shift, graph cut and MRF, to mention only a few. Contour-based image segmentation includes the active contour and active shape model, etc. All of these methods have their advantages for specific applications. During the course of our experiments each image contained multiple objects exhibiting relatively complex textures and intensity structures. None of the stan07/18/2006y introducing a level-set based deformable model which utilized Gaussian Hidden Markov Random Fields (GHMRF) improved the segmentation results significantly. In this new algorithm the level set function is propagated by the energy term defined on the likelihood map generated by GHMRF. In this way we can utilize the advantages of both approaches.
Results: The newly proposed level set coupled with GHMRF segmentation model provided accurate segmentation results for both synthetic images and for real imaged pathology specimens. One of the advantages of the algorithm is that it can operate on multiple cells or tissue-level image simultaneously, thus making parallel processing possible. An additional advantage to this approach is that it can be initialized with arbitrary starting positions.
Conclusions: To evaluate performance, we compared our results with several popular segmentation methods. Traditional gradient-based deformable models are not reliable due to the noise and background gradients of the imaged arrayss. The results of simple Gaussian Mixture models with EM algorithm failed to provide satisfactory results as it tended to give single background pixel settings inside the object, this is due to the lack of homogenous priors. Clustering based algorithm such as supervised Kth Nearest Neighbor, K-means or mean-shift tended to over-segment objects into non-homogenous parts, making it necessary to perform post-processing such as heuristic region merging. We are currently working to Grid-enable the proposed algorithm and test its performance using a much broader ensemble of specimens under a range of noisy conditions.