Date Submitted: Proposal Type: If renewal, current grant: Grant Application Resubmission? Prior App: RFA? RFA Title: TITLE OF PROJECT (Titles exceeding 81 characters, including spaces and punctuation, will be truncated.)

APPLICANT NAME HIGHEST DEGREE(S)

POSITION TITLE: APPLICANT’S CURRENT INSTITUTION ACADEMIC RANK: DIVISION: MAILING ADDRESS (Street, city, state, postal code, country)

DEPARTMENT:

E-MAIL ADDRESS:

Tel: Fax: PROGRAM ELIGIBILITY INFORMATION: (Responses to selected fields displayed below. For some grant programs this section may be blank.)

DATES OF PROPOSED PROJECT (MM/DD/YYYY) PROPOSED BUDGET ( ( From Through

SIGNING OFFICIAL FOR Name Name Address Title Address

Tel: Fax: Tel: Fax: EIN E-MAIL ADDRESS DUNS HUMAN SUBJECTS No Yes VERTEBRATE ANIMALS No Yes Human Subjects Assurance No. IRB Status: Animal welfare assurance no. IACUC Status: IRB Date: IACUC Date:

APPLICANT ASSURANCE: I certify that the statements herein are true, complete and SIGNATURE OF APPLICANT DATE accurate to the best of my knowledge. I am aware that any false, fictitious, or fraudulent (In ink. “Per” signature not acceptable.) statements or claims may subject me to criminal, civil, or administrative penalties. I agree to accept responsibility for the scientific conduct of the project and to provide the required progress reports if a grant is awarded as a result of this application. SIGNING OFFICIAL ASSURANCE: I certify that the statements herein are true, complete SIGNATURE OF SIGNING OFFICIAL DATE and accurate to the best of my knowledge, and accept the obligation to comply with the (In ink. “Per” signature not acceptable.) grantor’s terms and conditions if a grant is awarded as a result of this application. I am aware that any false, fictitious, or fraudulent statements or claims may subject me to criminal, civil, or administrative penalties. ADDITIONAL SIGNATURE (follow guidelines for required signatures): DATE ADDITIONAL SIGNATURE (follow guidelines for required DATE I certify that the statements herein are true, complete and accurate to signatures): I certify that the statements herein are true, the best of my knowledge. complete and accurate to the best of my knowledge.

Applicant: Application Contacts Role Role Name Name Institution Institution Title Title Division Division Dept Dept Address Address

Tel: Fax: Tel: Fax: E-mail E-mail Role Role Name Name Institution Institution Title Title Division Division Dept Dept Address Address

Tel: Fax: Tel: Fax: E-mail E-mail Role Role Name Name Institution Institution Title Title Division Division Dept Dept Address Address

Tel: Fax: Tel: Fax: E-mail E-mail Role Role Name Name Institution Institution Title Title Division Division Dept Dept Address Address

Tel: Fax: Tel: Fax: E-mail E-mail

GENERAL AUDIENCE SUMMARY

APPLICANT NAME DATE SUBMITTED

TITLE OF PROJECT (Titles exceeding 81 characters, including spaces and punctuation, will be truncated.)

This General Audience Summary will become public information; therefore, do not include proprietary/confidential information.

SCIENTIFIC ABSTRACT

APPLICANT NAME DATE SUBMITTED

TITLE OF PROJECT (Titles exceeding 81 characters, including spaces and punctuation, will be truncated.)

This Scientific Abstract will become public information; therefore, do not include proprietary/confidential information.

A. Title of Research Project Multiscale Shape/Intensity Distributions and Machine Learning for Automated Breast Cancer Histologic Grade Estimation B. General Audience Summary (Lay Abstract) Breast cancer is the most common malignant tumor in women, affecting approximately 1 in 9 women in the United States during their lifetime. Although substantial effort has been made both in the diagnosis and treatment of this disease, one third of breast cancer patients are ex- pected to die from it. Breast cancer is a progressive disease at all of its stages, and early detec- tion, accurate diagnosis and timely treatment can alter the natural course of the disease. Once a tumor is detected, a biopsy sample is taken, and studied by a pathologist (histology) to assess the stage of the disease. Correctly determining the development stage of the tumor will allow doc- tors to prescribe the most effective treatments. Breast cancers are classified according to their histologic grade, which to date is one of the most reliable ways to gauge overall survival. His- tologic grade ranges from I to III, with a grade I having a 5-year survival probability of 95% and a grade III having 5-year survival probability of 50%. Unfortunately different pathologists can give different grades to the same breast cancer sample. This occurs because there are no objec- tive methods for measuring the cellular structures and patterns found in breast cancer tumors. These discrepancies may lead to improper therapy for a particular patient. For example, some patients in the ``better'' prognosis category could actually have an aggressive form of the disease, indicating that chemo and hormone treatments are either unneeded, ineffective, or insufficient in various breast cancer patients. Therefore, unbiased methods for histologic grading are needed in order to remedy possible problems in the current grading procedure and improve prognosis. We propose to develop novel computational technologies that may be used to automatically and objectively estimate the histologic grade of breast cancer tumors. Our approach to auto- mated histologic grading is based on three powerful computing technologies: image processing, shape analysis and machine learning. Our approach involves acquiring digital images of cross- sections of breast cancer tumor specimens. The cancer cells in the image will be identified with image processing. The shape and distribution of the cancer cells will be mathematically modeled using shape analysis. The mathematical model of cancer cell shapes may then be analyzed, via machine learning, to determine what shapes and cell distributions are unique to Grades I, II and III breast carcinomas. Once we have mathematically described the shapes associated with these different grades, we can now identify the grade of an unknown sample by comparing the shapes and distribution of its cancer cells with the pre-classified identifiers. This computational ap- proach provides a standard method for histologic grading that will lead to a more reliable and consistent assessment of breast cancer and ultimately to improved cancer health care delivery. C. Scientific Abstract Background. Breast cancers can be categorized histologically based upon their architectural pat- terns and cellular types. Histologic grading has problems with inter- and intra-observer variabil- ity mainly due to a lack of an objective method for measuring architectural and cellular parame- ters of tumors. The variability is a product of both observer variation and sampling bias. Inaccu- rate histologic characterization can result in inappropriate treatment for a given patient. Compu- tational analysis of breast cancers offers an operator-independent method for histologic grading that should enhance grading reliability. Objective/Hypothesis. We propose to develop novel computational technologies that may be used to automatically and objectively estimate the histologic grade of breast cancer tumors. Our

approach to automated histologic grading is based on image processing and shape analysis of imaged histologic sections at multiple physical scales. Our hypothesis is that the structure of the cellular pleomorphisms found in breast cancer tumors can be transformed into distinct high- dimensional shape distributions using geometric measures from stochastic geometry. The result- ing shape distributions will map into well-separated regions of the high-dimensional space de- fined by the distributions. We will augment this space with information derived from the inten- sity variations of the hyperchromatism found in the cancer cell nuclei. Mapping an unknown breast cancer sample into this high-D space and determining, via machine learning, to which re- gion it belongs will allow us to automatically estimate its histologic grade. Specific Aims. The specific aims of this project focus on investigating and evaluating the image processing, shape analysis and machine learning technologies needed to develop a reproducible standard for automated estimation of breast cancer histologic grade. Study Design. We will develop a computational pipeline consisting of 1) accurate and robust automated cancer cell segmentation techniques that use advanced pattern clustering and recogni- tion techniques, 2) a variety of geometric measures from stochastic geometry that transform can- cer cell shapes into shape distributions, 3) creating intensity distributions from the texture varia- tion / nuclear hyperchromatism levels within cancer cells, 4) mapping the shape/intensity distri- butions into a high-dimensional space for separation analysis, 5) machine learning technologies (e.g. Support Vector Machines) that identify which geometric measures and intensity features will provide the most reliable/significant grade classifier, and 6) validating tests that utilize specimens from a breast cancer databank. Potential Outcomes and Benefits. The project will provide an initial implementation of a compu- tational, objective, reproducible method for determining the histologic grade of breast cancer tu- mors. Consistent and effective grading of tumors should provide improved health care delivery by indicating proper treatments for breast cancer patients. D. Background Once a breast cancer tumor is detected, a biopsy sample is taken, and studied histologically to assess its biological behavior by determining its histological type, histological grade, overall size and margins. Tumor typing correlates with prognosis and is a key factor in determining chemo and hormone therapy for a patient. Breast cancers can be typed histologically based upon their architectural patterns and cellular types. Histologic grades to date provide one of the most reli- able ways to gauge overall survival. The grades range from I to III, with a grade I having a 5- year survival probability of 95% and a grade III having 5-year survival probability of 50%. Histologic grading has problems with inter- and intra-observer variability mainly because of a lack of an objective method for measuring architectural and cellular parameters of tumors. This is compounded by the multitude of histologic types. When histologic grading is done in special- ized centers a good stratification of patients according to overall survival is obtained. This has not been the case when studies have looked at histologic grading performed at non-specialized health centers; thus highlighting the need for objective methods that will improve the reproduci- bility of the procedure [1,2]. Pathologist-based histologic grading is subject to observer variation on a given specimen sec- tion and the spatial focus of observation (sampling bias), and can be significant especially if the entire tumor is not studied histologically. Significant variations between different pathologists when grading histology samples can therefore occur. This may lead to improper therapy for a particular patient. For example, some patients in the ``better'' prognosis category will manifest aggressive disease and vice versa, indicating that chemo and hormone treatment is either un-

needed, ineffective, or insufficient in various breast cancer patients. Therefore, unbiased meth- ods for histologic grading are needed in order to remedy possible discrepancies in the current grading procedure. Developing a computational approach as a reproducible standard for his- tologic grading based on observed and quantifiable features of the samples will lead to a more reliable and consistent methodology for grading breast cancer tumors, which will ultimately pro- vide improved cancer health care delivery. E. Objective/Hypothesis We propose to develop novel computational technologies that may be used to automatically and objectively estimate the histologic grade of breast cancer tumors. Our approach to automated histologic grading is based on image processing and shape analysis of imaged histologic sections at multiple physical scales. Our hypothesis is that the structure of the cellular pleomorphisms found in breast cancer tumors can be transformed into distinct high-dimensional shape distribu- tions using geometric measures from stochastic geometry. Shape distributions [3] represent the structure/shape of 2D/3D objects with a probability distribution produced by randomly applying numerous geometric measures. Stochastic geometry is the study of the random processes that produce geometric structures and spatial patterns. It focuses on analyzing and quantifying the connections between geometry and probability in order to describe and characterize small-scale structural features, large-scale spatial events, and aggregate statistical geometric properties [4,5,6]. Specifically we propose to determine if shape distributions can uniquely capture and characterize the spatial distribution of neoplastic cells in breast cancer tumors. We hypothesize that, once the structure of the cells has been transformed, the resulting shape distributions of Grade I, II and III cancer cells will map into well-separated regions of the high-dimensional space defined by the distributions. We will augment this space with information derived from the intensity variations of the hyperchromatism found in the cancer cell nuclei. Mapping an un- known breast cancer sample into this high-D space and determining, via machine learning [7], to which region it belongs will allow us to automatically estimate its histologic grade.

Figure 1: Computational pipeline for automated histologic grade estimation. F. Specific Aims The specific aims of this project focus on investigating and evaluating the image processing, shape analysis and machine learning technologies needed to develop a reproducible standard for automated estimation of breast cancer histologic grade. The technologies can be placed in a computational pipeline, as seen in Figure 1, that begins with histology slides that are digitized at high magnification to produce high resolution images of the identified regions of interest. The pixel regions that contain cancer cells are identified to produce binary segmentation images. The

segmented images are analyzed to produce shape distributions that capture the spatial distribu- tion of the cell. The normalized intensity values in the original cancer cell pixels may also be transformed into a distribution. Each of these distributions can be interpreted as a single point in a high-dimensional space. Pre-graded breast cancer tumor samples will be processed, analyzed and mapped into the high-D shape distribution space. Populating the space with points from multiple samples will allow us to identify, via machine learning, those regions associated with different histologic grades. Once those regions have been identified, an unknown sample may be processed, mapped into the space, and its grade estimated. F.1 Specific Aim 1 – Automatic Segmentation of Cancer Cells in Areas of Interest Given H&E (Hematoxylin & Eosin) stained histology slides of breast cancer tumors develop ac- quisition and image processing techniques that will automatically digitize entire specimens, tar- get well-populated cell regions and re-capture those regions with high magnification, to produce images of regions of interest for further analysis. Develop techniques for automatically identify- ing and segmenting pixel regions in the images composed of cancer cells. F.2 Specific Aim 2 – Geometric Measures for Shape Distribution Transformation Develop and implement a variety of geometric measures that may be used to transform imaged cancer cell shapes into shape distributions. These geometric measures, also known as shape functions, will be drawn from stochastic geometry. They include the radial, line and triangle contact functions, as well as boundary distance functions. F.3 Specific Aim 3 – Generation of Cancer Cell Nuclei Intensity Distributions Develop techniques that create intensity distributions from the raw input image pixels corre- sponding to the segmentation pixels that identify cancer cell nuclei. The normalized intensities of these pixels will be mapped into distributions for each sample to represent texture variation / nu- clear hyperchromatism levels within cancer cells. F.4 Specific Aim 4 – Analysis of Distribution Space Numerous stained, digitized and pre-graded tissue samples of breast cancer tumors at different stages of the disease are available from a databank. The sample images will be processed and analyzed with our approach. The high-D points associated with the shape/intensity distributions produced from each sample will be grouped by histologic grade. Statistics will be gathered for each set of high-D points, for example the centroid, the standard deviation of the distances from the centroid, etc. These statistics will identify the region in the shape/intensity space occupied by each histologic grade for each mapping associated with a geometric measure; thus providing the information needed for histologic grade estimation. F.5 Specific Aim 5 – Machine Learning for Distribution Classification Investigate machine-learning technologies for determining which geometric measures and inten- sity features will best discriminate between the tumor grades/classes and provide the most reli- able/significant grade estimation classifier. Depending on how well these features discriminate between the classes, decision rules will be created to maximize certain criterion functions based on the computed distributions. F.6 Specific Aim 6 – Evaluation of Automated Histologic Grade Estimation Using the available breast cancer databank, numerous cases will be digitized, processed, and ana- lyzed and then used to validate the developed methods. Once the initial set of cases has provided enough training data to establish the detailed components of our histologic grade estimation ap- proach, tumors sample with unknown grade will be classified with our system. Experienced cancer pathologists will also grade these samples. The automated and the pathologist-produced grades will be compared in order to evaluate the automated approach.

Figure 2: Flowchart of cell-level automated processing of a breast cancer histology image. The stages include segmentation with adaptive thresholding and domain specific morphological op- erations, blob (micro-object) labeling, feature extraction and selection, and nuclei classification. G. Study Design G.1 Automatic Segmentation of Cancer Cells in Areas of Interest Breast cancer histology slides from our tumor databank will undergo the following computer automated image processing steps: a) the entire slide will be scanned using low power magnifi- cation; b) non-empty tissue areas within the slide will be identified and the coordinates of the smallest bounding boxes (BBox) surrounding these areas will be calculated; c) using the BBox coordinates the entire non-empty tissue areas will be scanned using intermediate power magnifi- cation; d) digital images created in c) will be processed using a hybrid segmentation method to identify high cell-density regions; e) calculated coordinates of these regions will be used to re- acquire these areas at high power magnification; f) the obtained high-resolution digital images of regions of high cell-density will be processed again to identify 10 different FOVs (field of views) with the highest cancer cell density; g) the 10 identified FOVs will be post-processed to optimize segmentation of the cancer cells within the FOV. The proposed automatic hierarchical digitization of slides will produce three sets of raw images and their corresponding segmenta- tions at different magnification powers. In preliminary breast tumor studies, we have developed an automated image-processing algo- rithm capable of detecting and identifying adipose tissue, extracellular matrix, morphologically distinct cell nuclei types, and the tubular organization of cells on the periphery of breast ducts using supervised learning [25]. Image processing and statistical analysis is composed of a hybrid segmentation method that combines adaptive thresholding with morphological operators, feature extraction, supervised learning, and non-linear classification. See Figure 2. This processing iden- tifies the spatial positions of hundreds of thousands of cell nuclei in three morphology groups in a single whole-section image. As a post-segmentation step, the resulting cell-level segmented images are further processed by analyzing the topological relation between individual neighbor- ing objects and their cell organization. The result is used to estimate the density distributions of the different cell morphologies identified earlier. The generated density distributions images are

then segmented and analyzed, using as reference the corresponding original color images and cell-level identification results, to detect higher order cell-structures within the tumor. As one of the aims of this project, we will increase the accuracy and robustness of our auto- mated segmentation by integrating advanced pattern clustering and recognition techniques. Within this module we will investigate unsupervised/semi-unsupervised clustering algorithms. A set of well-known unsupervised clustering algorithms will be tested from the vast collection of available clustering methods [8]. Among those to be considered will be: hierarchical fuzzy clus- tering, fuzzy k-means, and nearest neighbor in both agglomerative (merging clusters) and divi- sive (splitting cluster) algorithmic structure [8,9,10,11]. Of those investigated, the fittest subset of algorithms will be considered to design a novel hybrid clustering method. To summarize, this module will process raw breast tumor histological images and generate as output segmented im- ages to be used as input for stochastic analysis in the following aims.

Figure 3: A collection of shape distribution measures. D2: distance between two points on boundary. A3: angle formed by three points on the boundary. C1: spherical contact distance. C2: line contact length. C3: triangle contact area. Note that the contact distribution functions can measure the regions inside (pink) and outside (blue) the object of interest. e.g. C1.

G.2 Geometric Measures for Shape Distribution Transformation A shape distribution [3] is one type of digital signature that captures the shape and structure of 2D and 3D objects. It represents the features of an object as a probability distribution sampled from a shape function that measures geometric properties. Shape distribution analysis transforms a geometric object into a parameterized function of a geometric measure, e.g. distance, with a random sampling process providing the transformation. The information of the distribution func- tion is frequently represented as a histogram or Probability Density Function (PDF). Shape dis- tributions have been used extensively by the CAD and shape modeling communities and have shown to be effective when comparing and measuring the differences between the shapes of two objects [2,12]. Converting complex geometric models into a shape distribution maps the problem of comparing two complex and detailed models into the problem of measuring the distance be- tween two distributions. The latter is a well-studied problem with robust solutions, e.g LN norms and Earth Mover’s Distance, etc. [13,14,15,16]. Another closely related field that will contribute shape functions/measures to the proposed work is stochastic geometry. Stochastic geometry is the study of the random processes that produce geometric structures and spatial patterns. It fo- cuses on analyzing and understanding the connections between geometry and probability in order to describe and characterize geometric small-scale features and large-scale spatial events [4,5,6]. Evaluating shape functions (geometric measures) with a large set of randomly generated points on, in and/or around the model/region creates a shape distribution that capture the large-scale

shape of the model or region [23,24]. Shape functions are generally characterized by the number of points needed to calculate the geometric quantities, e.g. 1-point, 2-point, 3-point, etc. Osada et al. [2] defined several shape functions, a number of which may be useful in our work, namely D1 (distance between the shape centroid and a point on the shape's boundary), D2 (distance be- tween two points on the shape's boundary) and A3 (angle between three points on the shape's boundary). The D2 measure may be enhanced with multi-dimensional IN/OUT information to provide more accurate classification of surface models [12]. A number of contact distribution functions from stochastic geometry will also be investi- gated. It has been shown that the spherical (radial in 2D) contact distribution function C1 is ca- pable of capturing certain biological structures [17] and representing stochastic texture maps [18]. Another useful contact distribution function from stochastic geometry is the line contact distribution function C2, i.e. a type of fiber process [4,6]. This is a 2-point function that pro- duces a distribution consisting of the probabilities that a line of length l can be placed in the model without contact. A triangle contact distribution function C3 may also be defined. This is a 3-point function that produces a distribution consisting of the probabilities that a triangle of area A can be placed in the model without contact. Figure 3 illustrates the shape measures that will be initially investigated.

Figure 4: (left) Image of a Grade I breast cancer tumor (H&E, 20X Original Magnification). (right) Image of a Grade III breast cancer tumor (H&E, 20X Original Magnification).

G.3 Generation of Cancer Cell Nuclei Intensity Distributions Hyperchromatism increased nucleoli within cancer cell nuclei are one of the indicators for ele- vated histologic grade [19]. We will therefore isolate cancer cell nuclei and analyze the varia- tions of color intensities within them. The techniques described in Section G.1 will allow us to identify and segment those pixel regions in the sample images that contain cancer cells. The generated segmentation images may be used as masks that establish which pixels should be processed. The intensity values in these pixels will be evaluated and the number of each pixel with a particular intensity value will be tallied. This process will produce an intensity distribu- tion for each sample. As seen in Figure 4 the staining of cancer cells varies with histologic grade. The sample on the left is a Grade I. The staining in its nuclei is approximately uniform, and the intensity distribution derived from this sample will have a uni-modal form. The sample on the right is a Grade III, with nucleoli having clearly formed in the nuclei. The intensity distribution derived from this sample will have a bi-modal shape. These two different forms of intensity dis- tributions provide additional information that will be combined with the shape distributions to assist with the differentiation of the samples.

Figure 5: Possible locations of points in distribution space. Red points – distributions produced by applying the C1 measure to Grade I samples. Black point – distributions produced by apply- ing the D2 measure to Grade I samples. Blue points - distributions produced by applying the C1 measure to Grade III samples. In this (hypothetical) case the C1 measure produces well- separated clusters in distribution space.

G.4 Analysis of Distribution Space Shape and intensity distributions will be represented with histograms. Since a consistent method will be used to acquire the images (e.g. standard magnifications across samples), pixels may be used as the unit of length for the geometric measurements and the indices for the histogram bins. Tissue sample preparation will also be standardized, as well as the lighting and imaging condi- tions when acquiring sample images. This should lead to images with consistent intensity values across samples, allowing the histograms to use integer intensity values as bin indices. Given the consistent bin definition for the distribution histograms, the histograms may also be interpreted as points in a high dimensional space, where the dimension of the space is equal to the maximum number of bins used by the histograms. For example, if the shape distribution histogram uses 30 bins and the intensity distribution histogram uses 16 bins (due to intensity quantization), together they define a 46-dimension distribution space. The value assigned to each bin can be interpreted as the coordinate for the dimension associated with the bin. As tissue samples are processed and shape/intensity distributions are produced, the resulting high-D will be mapped into distribution space. Processing/analyzing more and more samples from the breast cancer databank will popu- late the distribution space with numerous data points. We will conduct a systematic application of the various geometric measures to the numerous breast cancer tumor samples available to us. Each transformation produced by the individual measures will produce a unique shape distribution, and therefore a point in distribution space. As we fill the distribution space with data points from the numerous samples and various geometric measures, we expect the data points to form identifiable and separable clusters in the high- dimensional space. See Figure 5 for a low-dimensional hypothetical example. Here, red points represent shape distributions generated with the C1 geometric measure from Grade I samples. The black points represent shape distributions generated with the D2 geometric measure from Grade I samples. The blue points represent shape distributions generated with the C1 geometric measure from Grade III samples. The x’s represent the centroids of the three clusters. It can

been seen that the red and blue points are well-separated, indicating that the C1 measure could provide reliable grade estimation. Since the black points overlap the blue ones, D2 may not be the optimal measure for grade estimation. The separation of the clusters will be initially be de- termined by analysis of the distance between the cluster centroids and the standard deviations of the distances from cluster points to the cluster centroid. Those measures that produce clusters with maximum separation will be further tested and evaluated in the final stages of the project. G.5 Machine Learning for Distribution Classification Given the shape and intensity distributions as characteristic features representing each tumor case, we will explore advanced machine learning techniques [7] that will automatically classify a specimen’s histologic grade. The classifications will be reproducible and consistent with the his- tologic grade assigned to each breast tumor case by pathologists. Using both labeled and unla- beled training data (shape and intensity distributions linked to a histologic grade provided by a team of pathologists) we will apply the following machine learning related tasks: 1) supervised and unsupervised feature selection (Forward and Backward search using N-fold Cross Valida- tion, K-Nearest Neighbor, Principal Component Analysis) to eliminate redundant/non- discriminative distributions, 2) unsupervised training/clustering (K-Means Clustering, Expecta- tion and Maximization Clustering) to identify possible predictive patterns, 3) supervised training (Binary Tree, Support Vector Machine, Quadratic Gaussian) that designs a histologic grade- estimation classifier. In the final step we will compare and analyze patterns identified by unsu- pervised clustering with classes resulting when using the supervised classifier. The study of Support Vector Machines (SVMs) as proposed by Vapnik [20,21] with non- linear kernel functions will be given top priority. We expect that the different shape functions will emphasize and capture different aspects of the input specimens. Therefore each measure should be weighted appropriately during classification. SVMs possess this desirable weighting property, where weights for the various measures are learned through an explicit training process and a maximum margin classifier is found from example data of different classes. Classifications of the input data based on histologic grade can be learned by feeding example shape descriptors (sets of dissimilarity measures) to SVMs. Learning the non-linear separating margins can be accomplished by using a radial basis kernel function [22], which projects the shape descriptors into a high dimensional feature space for the algorithm to find a separating margin. This results in a non-linear separating margin in the low dimensional input space. A significant research challenge here is the determination of the proper mapping, from the multi-dimensional shape descriptors into the vector space, that will be used to train the SVM. Additionally defining the most effective basis kernel function for categorizing biological models must also be addressed. G.6 Evaluation of Automated Histologic Grade Estimation We have access to a breast cancer databank that contains approximately 2200 paraffin-embedded breast cancer specimens collected from 1997 to the present. Each case record includes histologic classification, de-identified demographic data, and information on several prognostic markers including estrogen and progesterone receptors. All breast specimens have been submitted en- tirely using a standardized method. This protocol includes serial sectioning of the specimen at fixed intervals with consecutive ordering and maintenance of tissue orientation. Paraffin embed- ded 4µ thick hematoxylin and eosin stained sections of each block have been created and micro- scopically reviewed by multiple pathologists to allow classification of the tumor. Also, for each tumor case stained sections with estrogen receptor, progesterone receptor, Ki-67, p53 and Her2 breast cancer immunohistochemical biomarkers have been created. The database includes all types of breast cancer.

By the end of the project we expect to have a statistically significant number of cases for each grade imaged at multiple magnification powers and segmented. Part of the samples (equally representing each grade and benign cases) will be processed and analyzed, with their resulting shape/intensity distributions being mapped into distribution space. These will be the samples used in the Distribution Space Analysis study described in Section G.4. The remaining samples will be also be processed and analyzed. Their distributions will be mapped into distribu- tion space. We will calculate how close the “unknown” sample’s distribution space point is the Grade I, II and III distribution point clusters using the methods developed in Sections G.4 and G.5; thus estimating its grade. We expect that these tests will support our hypothesis that shape/intensity distributions may be used to estimate histologic grade in breast cancer tumors. H. Potential Outcomes The proposed project will develop an objective, reproducible computational method and proto- type software for automated estimation of breast cancer histologic grade. The evaluation of the method should provide evidence that highlights the capabilities of the method. This evidence will be the preliminary results needed to justify future clinical studies that will validate the effec- tiveness of the method. We will describe the method via numerous presentations and papers. Consistent and effective grading of tumors should provide improved health care delivery by en- suring proper and appropriate treatments are given to breast cancer patients. I. Timeline

Year 1 Year 2

Automatic Segmentation

Tissue Sample Processing

Geometric Measures

Intensity Distribution Generation

Distribution Space Analysis

Machine Learning

Method Evaluation

Dark bars – Heavy activity. Light bars – Light activity. J. Dissemination Plan The results of our research will be presented at a number of medical imaging and image analysis conferences, for example SPIE Medical Imaging, the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), and the IEEE Visualization Confer- ence. Papers describing our research will be published in the Breast Journal, IEEE Transactions on Medical Imaging and Medical Image Analysis. Additionally discussions have begun with one of the NIH National Centers for Biomedical Computing about including our software in their software toolkits for dissemination to the medical imaging community.

K. References [1] L. Dalton, S. Pinder, C. Elston, I. Ellis, D. Page, W. Dupont, and R. Blamey. Histologic grad- ing of breast cancer: Linkage of patient outcome with level of pathologist agreement. Modern Pathology, 13:730–735, 2000.

[2] C. Roberts, P. Beitsch, C. Litz, D. Hilton, G. Ewing, E. Clifford, W. Taylor, M. Hapke, A. Babaian, I. Khalid, J. Hall, G. Lindberg, K. Molberg, and H. Saboorian. Interpretive disparity among pathologists in breast sentinel lymph node evaluation. American Journal of Surgery, 186(4):324–329, 2003.

[3] R. Osada, T. Funkhouser, B. Chazelle, and D. Dobkin. Shape distributions. ACM Transac- tions on Graphics, 21(4):807–832, Oct. 2002.

[4] O. Barndorff-Nielsen, W. Kendall, and M. van Lieshout. Stochastic Geometry: Likelihood and Computation. Chapman & Hall, 1999.

[5] D. Stoyan, W. Kendall, and J. Mecke. Stochastic Geometry and Its Applications. John Wiley & Sons, 1987.

[6] V. Benes and J. Rataj, Stochastic Geometry: Selected Topics, Kluwer, 2004.

[7] T. Mitchell. Machine Learning. McGraw-Hill, 1997.

[8] A.K. Jain, M.N. Murty, and P.J. Flynn. Data Clustering: A Review. ACM Computing Sur- veys, Vol. 31(3) 1999

[9] A.B. Geva. Hierarchical Unsupervised Fuzzy Clustering. IEEE Transactions on Fuzzy Sys- tems, Vol. 7(6) 1999

[10] J.C. Bezdek. Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press, New York 1981

[11] R.N. Dave. Generalized Fuzzy c-Shell clustering and detection of circular and elliptical boundaries. Pattern Recognition, Vol. 25 (7) 1992

[12] C. Ip, D. Lapadat, L. Sieger, and W. Regli. Using shape distributions to compare solid mod- els. In Proc. Symposium on Solid Modeling and Applications, pages 273–280, June 2002.

[13] A. Bhattacharyya. On a measure of divergence between two statistical populations defined by their probability distributions. Bulletin of the Calcutta Mathematics Society, 35:99–110, 1943.

[14] J. Puzicha, Y. Rubner, C. Tomasi, and J. Buhmann. Empirical evaluation of dissimilarity measures for color and texture. In Proc. IEEE International Conference on Computer Vision, pages 1165–1173, 1999.

[15] Y. Rubner, C. Tomasi, and L. Guibas. The earth mover’s distance as a metric for image re- trieval. International Journal of Computer Vision, 40(2):99–121, 2000.

[16] M. Werman, S. Peleg, and A. Rosenfeld. A distance metric for multi-dimensional histo- grams. Computer, Vision, Graphics, and Image Processing, 32:328–336, 1985.

[17] C. Schroeder, W. Regli, A. Shokoufandeh, and W. Sun. Computer-aided design of porous artifacts. Computer-Aided Design, 37(3):339–353, March 2005.

[18] C. Schroeder, XXXXX, C. Cera, and W. Regli. Stochastic microgeometry for displacement mapping. In Proc. Shape Modeling International Conference, pages 164–173, 2005.

[19] H. Bloom and W. Richardson. Histological grading and prognosis in breast cancer. British Journal of Cancer, 1957;11:359–77.

[20] C. Cortes and V. Vapnik. Support-vector networks. Machine Learning, 20(3):273–297, 1995.

[21] C. J. C. Burges. A tutorial on support vector machines for pattern recognition. Data Mining and Knowledge Discovery, 2(2):121–167, 1998.

[22] C. Ip and W. Regli. Manufacturing classification of CAD models using curvature and SVMs. In Proc. Int. Conference on Solid Modeling and Applications, 2005.

[23] H. Robbins. On the measure of a random set I. Annals of Mathematical Statistics, 15:70–74, 1944.

[24] H. Robbins. On the measure of a random set II. Annals of Mathematical Statistics, 16:342– 347, 1945.

[25] XXXXX, XXXXX, XXXXX, C. Katsinis, and A. Tozeren. Large scale computations on histology images reveal grade-differentiating parameters for breast cancer. in review, BMC Medical Imaging, 2006.

BUDGET JUSTIFICATION Do NOT include any identifying information about the applicant. Do NOT exceed 2 Pages.

Personnel

The PI will devote half-time during the summer to the project (Year 1 - $15,435; Year 2 - $16,207) and 10% during the school year (no charge) for a total percentage of effort of 20%. He will provide geometric analysis/processing expertise for the Geometric Measures, Intensity Distributions, Distribution Space Analysis and Method Evaluation Tasks.

Collaborator 1 will devote 10% of his time to the project during the year, but will only charge for 4% (Year 1 - $5,036; Year 2 - $5,288). He will provide histologic grade estimation expertise for the Automatic Segmentation, Tissue Sample Processing and Method Evaluation Tasks and some project management.

Collaborator 2 will devote 5% of her time to the project during the year, but will only charge for 2% (Year 1 - $2,518; Year 2 - $2,644). She will provide secondary, supporting histologic grade estimation expertise for the Automatic Segmentation, Tissue Sample Processing and Method Evaluation Tasks.

A Research Scientist will devote half-time during the year ($32,00). He will provide image processing and machine learning expertise for the Automatic Segmentation, Tissue Sample Processing, Intensity Distributions, Machine Learning and Method Evaluation Tasks.

The benefits charge for the PI and the Research Scientist is 26.8% (Year 1 - $12,713; Year 2 - $13,348). The benefits charge for the two Collaborators is 14.4% (Year 1 – $1, 088; Year 2 - $1,142).

A graduate student will devote full-time to the project (Year 1 - $19,200; Year 2 - $20,160). He will contribute to the Geometric Measures, Intensity Distributions, Distribution Space Analysis and Method Evaluation Tasks.

Supplies – Storage media, software, computer supplies and lab supplies (Year 1 - $2,500; Year 2 - $2,500).

Equipment – Two high-end computers with high resolution monitors and 1TB of disk space for the research scientist and graduate student ($6,000 each) (Year 1 - $12,000).

Travel – One international conference (MICCAI) and two domestic conferences (IEEE Visualization and SPIE Medical Imaging) per year (Year 1 - $5,500; Year 2 - $5,500).

Other expenses – Graduate student tuition (Year 1 - $21,100; Year 2 - $21,100).

BIOGRAPHICAL SKETCH Provide the following information for the key personnel in the order listed on Form Page 2. Follow this format for each person. DO NOT EXCEED FOUR PAGES.

NAME POSITION TITLE Fernando U. Garcia, M.D. Professor of Pathology EDUCATION/TRAINING (Begin with baccalaureate or other initial professional education, such as nursing, and include postdoctoral training.) DEGREE INSTITUTION AND LOCATION YEAR(s) FIELD OF STUDY (if applicable) Universidad Peruana Cayetano Heredia MD 1981 Medicine University of Kansas, Kansas City, Kansas 1982-1986 Resident in Pathology Vanderbilt University, Nashville, Tennessee 1986-1987 Fellow in Surg Pathology Vanderbilt University, Nashville, Tennessee 1987-1988 Chief Resident Vanderbilt University, Nashville, Tennessee 1988-2000 Instructor in Pathology

A. Professional Experience:

1987-1990 Instructor, Department of Pathology, Vanderbilt University, Nashville, Tennessee. 1988-1990 Consultant, Laboratory Services, Veterans Administration Medical Center, Nashville, Tennessee. 1990-1991 Clinical Instructor, Pathology Department Kansas University, Kansas City, Kansas. 1990-1994 Director, Tissue Culture Core Facility, Pathology Department, Kansas University Medical Center, Kansas City, Kansas. 1991-1994 Medical Director, Cytopathology Laboratory Pathology Department, Kansas University, Kansas City, Kansas. 1991-1994 Assistant Professor, Pathology Department Kansas University, Kansas City, Kansas. 1992-1994 Consultant, Laboratory Services, Veterans Administration Medical Center, Leavenworth, Kansas. 1992-1994 Director of Dermatopathology, Kansas University, Kansas City, Kansas. 1994-1994 Director of Image Analysis, Kansas University, Kansas City, Kansas. 1995-2000 Associate Professor MCP-Hahnemann University Philadelphia, Pennsylvania. 1997-2001 Medical Director, Immunopathology Laboratory, MCP-Hahnemann University, Hahnemann University Hospital.

2001- 2005 Medical Director, Immunopathology ,Department of Pathology, Drexel College of Medicine Graduate Hospital, Philadelphia. 1997- 2001 Medical Director, Image Analysis Laboratory, MCP-Hahnemann University, Hahnemann University Hospital. 2001- Present Medical Director, Image Analysis Laboratory, Department of Pathology, Drexel College of Medicine, Graduate Hospital, Philadelphia. 1997- Present Medical Director, Tissue Procurement Core Laboratory, Department of Pathology, Drexel College of Medicine, Philadelphia. 1998- Present Regional Director of Anatomic Pathology, Department of Pathology, Drexel College of Medicine, Philadelphia. 2000-Present Professor of Pathology, Drexel University College of Medicine Philadelphia, Pennsylvania.

Certifications: 1981 Peru Medical License (13087) 1981 Educational Commission for Foreign Medical Graduates Certificate (329-894-0) 1984 Visa Qualyfing Examination (VQE) 1985 Federal Licensure Examination (FLEX) 1987 Tennessee Medical License (MD 17629) Not Active 1987 Diplomate of the American Board of Pathology in Anatomic Pathology 1990 Kansas Medical License (4-23066) Not Active 1993 Diplomate of the American Board of Pathology in Cytopathology 1995 Pennsylvania Medical License (MD-054895-L)

B. Selected Publications

1. Stella, J.J., Bontempo Jr., D., Thomas, M.P., Garcia, F.U., Samuels, L.E. Castelman’s disease of the retroperitoneum: Case report and review of the literature. Surgical Rounds; 1999; 22(3):143-148. 2. Garcia, F.U. and Soans, S. Malignant Cystosarcoma phyllodes with sarcomatous overgrowth. Report of a case with cytologic presentation. Diagnostic Cytopathology; 1999; 20(4):241-245. 3. Soans, S., Galindo, L.M., Garcia, F.U. Mucin Stain on Frozen Sections: A rapid 3- minute method. Arch Pathol & Lab Med; 1999, 13(5):362-364. 4. Shidham, V.B., Ayala, G.E., Garcia, F.U. Low grade fibromyxoid sarcoma. A case report with review of the literature. Am J Clin Oncol; 1999, 22(2):150-155. 5. Wang, M., Liu, A., Garcia, F.U., Rhim, J.S., Stearns, M.E. Growth of HPV-18 immortalized human prostatic cell lines. Influence of IL-10, follistatin, activin-A, and DHT. Int J Oncol; 1999, 14(6):1185-1195. 6. Stearns, M.E., Garcia, F.U., Fudge, K.A., Wang, M. Role of interleukin 10 and transforming growth factor beta1 in the angiogenesis and metastasis of human prostate primary tumor lines from orthotopic implants in severe combined immunodeficiency mice. Clinical Cancer Research; 1999, 5(3):711-720. 7. Shidham, V.B., Lindholm, P., Kajdacsy-Balla, A., Basir, Z., George, V., Garcia, F.U. PSA expression and lipochrome pigment granules in the differential diagnosis of prostatic adenocarcinoma versus ejaculatory duct-seminal vesicular epithelium. Arch Pathol & Lab Med, 1999, Vol. 123(11):1093-1097. 8. Galindo, L.M., Garcia, F.U. Hanau, C., Lessin, S.R., Jhala, N., Bigler, R.D., Vonderheid, E. Fine needle aspiration biopsy in the evaluation of lymphadenopathy associated with cutaneous T-cell lymphoma (Mycosis fungoides/Sezary syndrome. Am J Clin Pathol, 2000, 113(6): 865-871. 9. Garcia, F.U., Taylor, C.A., Hou, J.S., Stearns, M.E. Rukstalis, D.B. Increased cellularity of prostate carcinoma encased native vessels is a marker for tumor progression. Modern Pathol, 2000, 13(7):717-722. 10. Rukstalis, D.B., Khorsandi, M., Garcia, F.U., Hoenig, D.M., Cohen, J. Clinical experience with open renal cryoablation. Urology, 2000, 57(1):34-39. 11. Galindo, L.M., Dwyer-Joyce, L., Shienbaum, A.J. and Garcia, F.U. Atypical hemangioma of the breast. A diagnostic Pitfall in breast fine needle aspiration. Diagnostic Cytopathology, 2000, 24(3): 215-218. 12. Leon, ME, Chavez, C, Fyfe, B, Nagosky, MJ, Garcia, F.U. Case Report: Cholesterol granuloma of the maxillary sinus. Arch Pathol & Lab Med 2002, 126(2):217-219. 13. Haber, MM, Lu, L, Garcia, F.U. Use of MIB-1 in the assessment of esophageal biopsies from patients with gastroesophageal reflux disease in well and poorly oriented areas. Applied Immunohistochemistry & Molecular Morphology, 2002, 10(2):128-133. 14. Giordano, A, Cardi, N, Paulo, P and Garcia, F.U. Expression of cell-cycle regulated proteins pRb2/p130, p107, p27ki1, p53, Mdm-2 and Ki-67 (MIB-1) in prostatic gland adenocarcinoma. Clinical Cancer Research, 2002, 8(6):1808-1815. 15. Leon, ME, Leon, MA, Ahuja, J, Garcia, F.U. Nodular fibroblastic stromal hyperplasia of the mammary gland as an accurate name for pseudoangiomatous hyperplasia of mammary stroma. The Breast Journal 2002, 8(5): 1-4. 16. Rukstalis, D.B., Goldknoff, J.L., Crowley, E.M. and Garcia, F.U. Prostate cryoablation: A scientific rationale for future modifications. Urology, 2002, 60 (Suppl 2A): 19-25. 17. Stearns, M.E., Wang, M. Hu, Y., Garcia, F.U. and Rhim, J. IL-10 blocks MMP-2 and MT1-MMP synthesis in primary human prostate tumor lines. Clin. Can. Res. In press. October, 2002. 18. Haber MM, Sharma M, Leon ME, Nagle DA, Garcia FU. Interobserver variability in grading of anal intraepithelial neoplasia is reduced by mib-1 immunostaining. USCAP 2004, Vancouver, BC, Canada* 19. Garcia FU, Selim SS, Leon RA, Popnikolov, NK, Haber MM. Basal cell cocktail (34ßE12 + p63) distinguishes zonal variation of basal cells and peripheral zone intermediate basal cell hyperplasia. USCAP 2004, Vancouver, BC, Canada* 20. Nagle D, Butcher D, Garcia FU, Haber MM, Sharma M. Increased sensitivity of anal cytology in evaluation of internal compared with external lesions. American Society of Colon and Rectal Surgeons, Dallas, May 2004. 21. S. Petushi, C. Katsinis, C. Coward, F. Garcia, and A. Tozeren. Automated identification of microstructures on histology slides. IEEE International Symposium in Biomedical Imaging, 2004 22. S. Petushi, F. Garcia, M. Haber, C. Katsinis, and A. Tozeren. Large scale computations on histology images reveal grade-differentiating parameters for breast cancer. BMC Medical Imaging, 2006. Under review

C. Research Support

Ongoing Research Support:

Title: Predictive Syndromic Surveillance System (PS3) Granting Agency: Army TATRC (DOD) Position: PI Dates: 6/2005 - 8/2007 Total Direct Costs: 400,000 Status: In Progress

Completed Research Support:

Grants: Title Project: University of Kansas Cancer Center Breast Tissue and Serum Repository Core Facility. Granting Agency: United States Army Medical Research and Development Command. Total Direct Costs: $404,369 Position: Co-Principal Investigator Dates: 5/1994-5/1998 Status: Completed.

Contract MAO/RFP No NCI-CN-750006-63 In vitro screening of potential chemopreventive agents using human prostate epithelial cells Workstatement # 58 Total Costs: $ 289,000 Position: Co-Principal Investigator Dates: 6/97-12/98 Status: Completed

Title Project: Role of IGF-R1and metalloproteinase-2 in prostate cancer Granting Agency: NIH-NCI Total Direct Costs: $189, 259 Percent Effort: 2 Position: Collaborator Dates: 10/98-10/01 Status: Completed

Title: Magnetic imaging of the breast: A non-invasive mean of distinguishing benign from malignant abnormalities. Granting Agency: Susan G. Komen Breast Cancer Foundation Total Direct Costs: $240,000 Position: Collaborator Dates: 3/99-3/02 Status: Completed

Title: Genetics origins of premature ovarian failure. Granting Agency: Institute for Womenís Health and The Institute for Humanities, Drexel University. Mini-Challenge Grant Application Total Direct Costs: $4,900/yr Position: Co-Principal Investigator Dates: 4/2000- 6/2001 Status: Completed

Title: A novel urine marker for prostate cancer Granting Agency: Department of Defense Total Direct Costs: $125,000/yr Position: Co-Principal Investigator Dates: 6/2000-2003 Status: Completed

Title: Over-expression of a PSTF-1 chromosomal protein is prognostic for prostate cancer. Granting Agency: NIH-NCI Total Direct Costs: $464,000/yr Position: Co-Principal Investigator Dates: 8/99-7/03 Status: Completed

BIOGRAPHICAL SKETCH Provide the following information for the key personnel in the order listed on Form Page 2. Follow this format for each person. DO NOT EXCEED FOUR PAGES.

NAME POSITION TITLE Marian Markowitz Haber, MD Professor of Pathology EDUCATION/TRAINING (Begin with baccalaureate or other initial professional education, such as nursing, and include DEGREE INSTITUTION AND LOCATION YEAR(s) FIELD OF STUDY (if applicable) Wellesley College, Wellesley, M.A. B.A. 1979-83 Psychobiology, Religion Columbia University, College of P & S, NY,NY M.D. 1983-87 Medicine Columbia Presbyterian Med Ctr, NY, NY 1987-91 Pathology residency Yale University 1991-93 Pathology fellowship

Professional Experience: 1987 - 1990 Resident - Department of Pathology, Columbia Presbyterian Medical Center, New York, New York 1990 - 1991 Chief Resident - Department of Pathology, Columbia Presbyterian Medical Center, New York, New York 1991 - 1993 Postdoctoral Fellow - Gastrointestinal Pathology, Yale University School of Medicine, New Haven, Connecticut 1993 - 1994 Assistant Professor of Pathology, Department of Pathology and Laboratory Medicine, Allegheny University Hospitals, MCP Division, Philadelphia, PA 1994 - 1998 Assistant Professor of Pathology, Department of Pathology and Laboratory Medicine, Allegheny University Hospitals, Hahnemann Division, Philadelphia, PA 1998 - 1999 Assistant Professor of Pathology, Department of Pathology and Laboratory Medicine, MCP Hahnemann University, Hahnemann Division, Philadelphia, PA 1999 - 2001 Associate Professor of Pathology, Department of Pathology and Laboratory Medicine, MCP Hahnemann University, Hahnemann Division, Philadelphia, PA 2001 - 2004 Associate Professor of Pathology, Department of Pathology and Laboratory Medicine, Drexel University College of Medicine, Graduate Hospital, Philadelphia, PA 2004 to present Professor of Pathology, Director of Pathology, Department of Pathology and Laboratory Medicine, Drexel University College of Medicine, Graduate Hospital, Philadelphia, PA

Certification: Diplomate of NBME #339009 Licentiate of Pennsylvania State #MD-050427L Diplomate of the American Board of Pathology #91-441

Honors and Awards: 1980 Freshman Distinction 1982 Phi Beta Kappa 1983 Magna Cum Laude 1987 Harold Lee Meierhof Memorial Prize for Pathology

Publications (last 3 years): 1. *Shidham V, Gupta D, Galindo L, Haber M, Grotkowki C, Edmonds P, Subichin SJ, George V, England J. Intra-operative scrape cytology: comparison with frozen section using receiver operating characteristic curve. Diagnostic Cytopathology 2000;23:134-139. 2. *Agrawal NM, Campbell DR, Safdi DR, Lukasik NL, Huang B, Haber MM. Superiority of lansoprazole vs ranitiding in healing nonsteroidal anti-inflammatory drug-associated gastric ulcers: results of a double-blind, rnadomized, mulitcenter study. NSAID-Associated Gastric Ulcer Study Group. Arch Intern Med 2000;160:1455-1461. 3. *Kozar R, Cipolla J, Haber M. Antioxidant enzymes are induced during recovery from acute lung injury. Critical Care Medicine 2000:28:2486-2491. 4. *Campbell DR, Haber MM, Sheldon E, Collis C, Lukasik N, Huang B, Goldstein JL. Effect of H. pylori Status on Gastric Ulcer Healing in Patients Continuing Nonsteroidal Anti-inflammatory Therapy and Receiving Treatment With Lansoprazole or Ranitidine. Am J Gastroenterol 2002:97:2208-14. 5. *Mohsen NA, Haber M, Urrutia VC, Nunes LW. Intimal sarcoma of the aorta. AJR Am J Roentgenol 2000;175:1289-90. 6. *Hirschowitz BI, Haber MM. Helicobacter pylori effects on gastritis, gastrin and enterochromaffin- like cells in Zollinger-Ellison syndrome and non-Zollinger-Ellison syndrome acid hypersecretors treated long-term with lansoprazole. Aliment Pharacol Ther. 2001;15:887-103. 7. *Keates J, Laghee S, Crilley P, Haber M, Kowalski T. CMV enteritis causing segmental ischemia and massive intestinal hemorrhage. Gastroeintest Endosc 2001;53:355-9. 8. *Montogmery E, Bronner MP, Goldblum JR, Greenson JK, Haber MM, Hart J, Lamps LW, Lauwers GY, Lazenby AJ, Lewin DN, Robert ME, Toledano AY, Shyr Y, Washinton K. Reproducibility of the diagnosis of dysplasia in Barretts esophagus: A reaffirmation. Hum Pathol 2001; 32:368-78. 9. *Montgomery E, Goldblum JR, Greenson JK, Haber MM, Lamsp LW, Lauwers GY, Lazenby AJ, Lewin DN, Robert ME, Washington K, Zahurak ML, Hart J. Dysplasia as a predictive marker for invasive carcinoma in Barrett esophagus: A follow-up study based on 138 cases from a diagnostic variability study. Hum Pathol 2001:32:379-88. 10. *Kozar RA, Kaplan LJ, Cipolla J, Meija J, Haber MM. Laparoscopic repair of traumatic diaphragmatic injuries. J Surg Res 2001;97:164-71. 11. *Graham DY, Agrawal NM, Campbell DR, Haber MM, Collis C, Lukasik NL, Huang B. Ulcer prevention in long-term users of nonsteroidal anti-inflammatory drugs: results of a double-blind, randomized, multicenter, active- and placebo-controlled study of misoprostol vs lansoprazole. Arch Intern Med. 2002 Jan 28;162(2):169-75. 12. *Montgomery E, Bronner MP, Greenson JK, Haber MM, Hart J, Lamps LW, Lauwers GY, Lazenby AJ, LewinDN, Robert ME, Washington K, Goldblum JR. Are ulcers a marker for invasive carcinoma in Barrett's esophagus? Data from a diagnostic variability study with clinical follow-up. Am J Gastroenterol. 2002 Jan;97(1):27-31. 13. *Haber MM, Lu L, Modi A, Garcia FU. Use of MIB-1 in the assessment of esophageal biopsy specimens from patients with gastroesophageal reflux disease in well- and poorly oriented areas. J Appl Immunohistochem Mol Morphol 2002;10:128-33. 14. Haber MM. Histologic precursors of gastrointestinal tract malignancy. Gastroenterol Clin N Am 2002; 31:395-419 15. *Kozar RA, Schultz SG, Hassoun HT, Desoignie R, Weisbrodt NW, Haber MM, Moore FA. The type of sodium-coupled solute modulates small bowel mucosal injury, transport function, and ATP after ischemia/reperfusion injury in rats. Gastroenterol 2002;123:8210-16. 16. Haber MM. Histologic precursors of gastrointestinal tract malignancy. Hematology and Oncology Clin N Am 2003; 17. Jhala NC, Montemor M, Jhala D, Lu L, Talley L, Haber MM, Lechago J. Pancreatic Acinar Cell Metaplasia in Autoimmune Gastritis.Arch Pathol Lab Med. 2003 Jul;127(7):854-857. 18. Haber MM, Leon ME, Bakker JE, Nagle D. Carcinoembryonic antigen elevation due to bowel sequestration with mucocele formation following colonic resection. Arch Pathol Lab Med 2003;127:1376-79.

G. BIOGRAPHICAL SKETCHES

BIOGRAPHICAL SKETCH Provide the following information for the key personnel in the order listed on Form Page 2. Follow this format for each person. DO NOT EXCEED FOUR PAGES.

NAME POSITION TITLE Sokol Petushi, M.Sc. Research Systems Engineer EDUCATION/TRAINING (Begin with baccalaureate or other initial professional education, such as nursing, and include postdoctoral training.) DEGREE INSTITUTION AND LOCATION YEAR(s) FIELD OF STUDY (if applicable) Technological Institute of Thessaloniki, Greece, EU B.A. 1998-02 Electronics Engineering Rochester Institute of Technology, Rochester, NY M.Sc. 2002-04 Computer Engineering Drexel University, Philadelphia, PA 2004-current Drexel University, Philadelphia, PA Ph.D. Biomedical Engineering Part-time

Professional Experience:

Aug 2005 - Current Research Engineer - Advanced Pathology Imaging Laboratory, Drexel University College of Medicine, Philadelphia, PA Jun 2004 - Jul 2005 Associate Research Engineer - Center for Integrated Bioinformatics, Drexel University, Philadelphia, PA Sep 2001 - Jul 2002 R&D Engineer (Traineeship ) - Industrial Vision Department, PHILIPS CFT, Eindhoven, The Netherlands, EU Mar 2001 - Aug 2001 Electronics Engineer (Co-Op)- R&D Department, COMPUCON S.A, Thessaloniki, Greece, EU

Scholarships and Honors

1. Research Assistantship at Drexel University (2003 - 2004) 2. Graduate Scholarship at Rochester Institute of Technology (2002) 3. Scholarship from the “Socrates” program of the European Union to complete the bachelor diploma thesis in The Netherlands (2001) 4. Honored to lead the Graduation Oath as the best student of the Electronics Department at Thessaloniki Institute of Technology (2001)

Publications 1. S. Petushi, F. Garcia, M. Haber, C. Katsinis, and A. Tozeren. Large scale computations on histology images reveal grade-differentiating parameters for breast cancer. BMC Medical Imaging, 2006. In review 2. S. Petushi, C. Katsinis, C. Coward, F. U. Garcia, and A. Tozeren. Automated identification of microstructures on histology slides. IEEE International Symposium in Biomedical Imaging, 2004 3. A. Tozeren, C. W. Coward, S. Petushi. Origins and Evolution of Cell Phenotypes in Breast Tumors. Journal of Theoretical Biology, Elsevier, 2004 4. S. Petushi, S. Efstratiadis and R. Wetzels. Visual Inspection of Semiconductor Crystals. 4th International Conference on Technology and Automation. Thessaloniki, Greece. October 2002 Applicant Name (Last, first, middle): BIOGRAPHICAL SKETCH Provide the following information for the applicant, PI and all key personnel. Please attach a separate biosketch for each person. Follow the sample format for each person found in Biosketch Sample. DO NOT EXCEED FOUR (4) PAGES.

NAME POSITION TITLE Breen, David Edward Assistant Professor

EDUCATION/TRAINING (Begin with baccalaureate or other initial professional education, such as nursing, and include postdoctoral training.)

INSTITUTION AND LOCATION DEGREE YEAR(s) FIELD OF STUDY (if applicable) Colgate University, Hamilton, NY BA 1982 Physics Rensselaer Polytechnic Institute, Troy, NY MS 1985 Computer & Systems Engineering Rensselaer Polytechnic Institute, Troy, NY PhD 1993 Computer & Systems Engineering

A. Positions and Honors.

Positions and Employment

1982-1985 Graduate Research Assistant, Center for Interactive Computer Graphics Rensselaer Polytechnic Institute, Troy, NY

1985-1993 Research Engineer, Visual Technologies Program, Design Research Center Rensselaer Polytechnic Institute, Troy, NY

1987-1988 Visiting Research Engineer, Computer Animation Group Fraunhofer Institute for Computer Graphics, Darmstadt, Germany

1994-1996 Member of Research Staff, User Interaction & Visualization Group European Computer-Industry Research Centre, Munich, Germany

1996-2002 Assistant Director, Computer Graphics Laboratory, Department of Computer Science California Institute of Technology, Pasadena, CA

2002-2003 Senior Research Scientist, Center for Advanced Computing Research California Institute of Technology, Pasadena, CA

2003-present Assistant Professor, Department of Computer Science Drexel University, Philadelphia, PA

Professional Memberships

1986-present Member, IEEE Computer Society 1986-present Member, Association for Computing Machinery 2000-present Member, Eurographics Association

Honors

1982 Awarded Garden State Graduate Fellowship

ALA Application Template 01/26/2004 Page ____ Biographical Sketch

Applicant Name (Last, first, middle): 1996 The Literati Club Award for Best Paper of the Year in the International Journal of Clothing Science and Technology (with D. House and R. DeVaul).

B. Selected peer-reviewed publications. (Selected from 55 peer-reviewed publications)

Book

1. D.H. House and D.E. Breen (eds.), Cloth Modeling and Animation, AK Peters, Natick, MA, 2000

Book Chapters

2. D. Breen, R. Whitaker, K. Museth and L. Zhukov, “Level Set Segmentation of Biological Volume Datasets,” J. Suri (ed.), Handbook of Medical Image Analysis, Volume I: Segmentation Part A, Kluwer, New York, Chapter 8, pp. 415-478, 2005 3. K. Museth, R.T. Whitaker, and D.E. Breen, “Editing Geometric Models,” S. Osher and N. Paragios (eds.), Geometric Level Set Methods in Imaging, Vision and Graphics, Springer, New York, Chapter 23, pp. 441-460, 2003 4. R.T. Whitaker, D.E. Breen, K. Museth and N. Soni, “Segmentation of Biological Volume Datasets Using a Level-Set Framework,” K. Mueller, A. Kaufman (eds.), Volume Graphics 2001, Springer, Vienna, pp. 249-263, 2001 5. D.E. Breen, S. Mauch and R.T Whitaker, “3D Scan Conversion of CSG Models into Distance, Closest-Point and Colour Volumes,” M. Chen, A.E. Kaufman, R. Yagel (eds.), Volume Graphics, Springer, London, Chapter 8, pp. 135-158, 2000

Journal Publications

6. I. Braude, J. Marker, K. Museth, J. Nissanov and D. Breen, “Contour-Based Surface Reconstruction using MPU Implicit Models,” to be published in Graphical Models, 2006 7. K. Museth, D.E. Breen, R.T. Whitaker, S. Mauch and D. Johnson, “Algorithms for Interactive Editing of Level Set Models,” Computer Graphics Forum, Vol. 24, No. 4, pp. 821-841, December 2005 8. L. Zhukov, K. Museth, D.E. Breen, R.T. Whitaker and A.H. Barr, “Level Set Segmentation and Modeling of DT-MRI Brain Data,” Journal of Electronic Imaging, Vol. 12, No. 1, pp. 125-133, January 2003 9. K. Museth, D.E. Breen, R.T. Whitaker and A.H. Barr, “Level Set Surface Editing Operators,” ACM Transactions on Graphics (Proc. SIGGRAPH 2002), Vol. 21, No. 3, pp. 330-338, July 2002 10. M. Aono, D.E. Breen and M.J. Wozny, “Modeling Methods for the Design of 3D Broadcloth Composite Parts,” Computer-Aided Design, Vol. 33, No. 13, pp. 989-1007, November 2001 11. D.E. Breen and R.T. Whitaker, “A Level-Set Approach for the Metamorphosis of Solid Models,” IEEE Transactions on Visualization and Computer Graphics, Vol. 7, No. 2, pp. 173-192, April- June 2001 12. D.E. Breen, “Cost Minimization for Animated Geometric Models in Computer Graphics,” Journal of Visualization and Computer Animation, Vol. 8, No. 4, pp. 201-220, 1997 13. G.J. Klinker, K.H. Ahlers, D.E. Breen, P.-Y. Chevalier, C. Crampton, D.S. Greer, D. Koller, A. Kramer, E. Rose, M. Tuceryan and R.T. Whitaker, “Confluence of Computer Vision and Interactive Graphics for Augmented Reality,” Presence: Teleoperations and Virtual Environments, Vol. 6, No. 4, pp. 433-451, August 1997 14. D.H. House, R.W. DeVaul and D.E. Breen, “Towards Simulating Cloth Dynamics Using Interacting Particles,” International Journal of Clothing Science and Technology, Vol. 8, No. 3, pp. 75-94, 1996 (Awarded Best Paper of the Year)

ALA Application Template 01/26/2004 Page ____ Biographical Sketch

Applicant Name (Last, first, middle): 15. M. Aono, P. Denti, D.E. Breen and M.J. Wozny, “Fitting a Woven Cloth Model to a Curved Surface: Dart Insertion,” IEEE Computer Graphics and Applications, Vol. 16, No. 5, pp. 60-70, September 1996 16. D.E. Breen, R.T. Whitaker, E. Rose and M. Tuceryan, “Interactive Occlusion and Automatic Object Placement for Augmented Reality,” Computer Graphics Forum (Proc. Eurographics ‘96), Vol. 15, No. 3, pp. 11-22, August 1996 17. M. Tuceryan, D.S. Greer, R.T. Whitaker, D.E. Breen, C. Crampton, E. Rose and K.H. Ahlers, “Calibration Requirements and Procedures for a Monitor-Based Augmented Reality System,” IEEE Transactions on Visualization and Computer Graphics, Vol. 1, No. 3, pp. 255-273, September 1995 18. D.E. Breen, D.H. House and M.J. Wozny, “A Particle-Based Model for Simulating the Draping Behavior of Woven Cloth,” Textile Research Journal, Vol. 64, No. 11, pp. 663-685, November 1994 19. M. Aono, D.E. Breen and M.J. Wozny, “Fitting a Woven Cloth Model to a Curved Surface: Mapping Algorithms,” Computer-Aided Design, Vol. 26, No. 4, pp. 278-292, April 1994

Refereed Conference Publications

20. J. Marker, I. Braude, K. Museth and D. Breen, “Contour-based Surface Reconstruction Using Implicit Curve Fitting, and Distance Field Filtering and Interpolation,” Proc. International Workshop on Volume Graphics, July 2006, pp. 95-102 21. O. Nilsson, D. Breen and K. Museth, “Surface Reconstruction Via Contour Metamorphosis: An Eulerian Approach With Lagrangian Particle Tracking,” Proc. IEEE Visualization 2005, pp. 407- 414 22. C. Schroeder, D. Breen, C. Cera and W. Regli, “Stochastic Microgeometry for Displacement Mapping,” Proc. International Conference on Shape Modeling and Applications, June 2005, pp. 164-173 23. K. Museth, D. Breen, L. Zhukov and R. Whitaker, “Level Set Segmentation From Multiple Non- uniform Volume Datasets,” Proc. IEEE Visualization 2002, pp. 179-186 24. M. Gavriliu, J. Carranza, D. Breen and A. Barr, “Extracting Multiresolution Meshes from Volumes with Guaranteed Properties,” Proc. IEEE Visualization 2001, pp. 295-302 25. S. Lombeyda, L. Moll, M. Shand, D. Breen and A. Heirich, “Scalable Interactive Ray-casting of Volumes Using Off-the-shelf Components,” Proc. 2001 Symposium on Parallel and Large-Data Visualization and Graphics, pp. 115-121 26. Z. Wood, M. Desbrun, P. Schröder and D. Breen, “Semi-Regular Mesh Extraction From Volumes,” Proc. IEEE Visualization 2000, pp. 275-282 27. D.E. Breen, S. Mauch and R.T. Whitaker, “3D Scan Conversion of CSG Models into Distance Volumes,” Proc. 1998 Symposium on Volume Visualization, pp. 7-14 28. R.T. Whitaker and D.E. Breen, “Level-Set Models for the Deformation of Solid Objects,” Proc. 3rd International Workshop on Implicit Surfaces, June 1998, pp. 19-35 29. D.E. Breen, D.H. House and M.J. Wozny, “Predicting the Drape of Woven Cloth Using Interacting Particles,” Proc. SIGGRAPH '94 Conference Proceedings, pp. 365-372 30. M. Aono, D.E. Breen and M.J. Wozny, “A Computer-Aided Broadcloth Composite Layout Design System,” Geometric Modeling for Product Realization (Proc. IFIP Conference on Geometric Modeling), September 1992, pp. 223-250 31. J.V. Miller, D.E. Breen, W.E. Lorensen, R.M. O'Bara and M.J. Wozny, “Geometrically Deformed Models: A Method for Extracting Closed Geometric Models from Volume Data,” Proc. SIGGRAPH '91 Conference, pp. 217-226 32. D.E. Breen, D.H. House and P.H. Getto, “A Particle-Based Computational Model of Cloth Draping Behavior,” Scientific Visualization of Physical Phenomena (Proc. CG International '91), pp. 113- 134 33. J.V. Miller, D.E. Breen and M.J. Wozny, “Extracting Geometric Models Through Constraint Minimization,” Proc. IEEE Visualization '90, pp. 74-82 ALA Application Template 01/26/2004 Page ____ Biographical Sketch

Applicant Name (Last, first, middle):

C. Research Support

Ongoing Research Support

D. Breen (PI) 2006-2007 NSF $75,000 Automated Shape Composition Via Genetic Programming. This project is exploring self-organizing geometric modeling primitives based on cell-based interactions and genetic programming.

F. Garcia (PI) 2005-2007 U.S. Army MRMC $135,000 Predictive Syndromic Surveillance System (Modeling Task). This project is developing digital information technology to improve disease diagnosis accuracy. Role: co-PI

M. Kam (PI) 2004-2007 U.S. Army CECOM $445,000 Modeling and Simulation of On-the-Move Networks (Visualization Task). This project is developing network visualization techniques. Role: co-PI

Completed Research Support

D. Breen (PI) 2004-2005 Drexel University $15,000 Contour-based Neuro-anatomical Surface Reconstruction: Application to Bioinformatics. This project developed a technique for reconstructing surfaces from delineated contours.

D. Breen (PI) 2000-2005 NSF $265,000 Interactive Level-Set Modeling for Visualization of Biological Volume Datasets. This project developed technologies needed for interactively specifying/modifying level set models.

P. Dimotakis (PI) 2000-2006 NSF $237,000 Development of the Distributed Teravoxel Data System: Acquisition, Networking, Archiving, Analysis and Visualization (Visualization Component). This project developed a system for interactively displaying large-scale volume datasets. Role: co-PI

D. Breen (PI) 1999-2003 NSF $1.2M Multiresolution Visualization Tools for Interactive Analysis of Large-Scale N-Dimensional Datasets. This was a multi-investigator project that produced techniques for volume data segmentation, model compression, multiresolution mesh extraction, tensor field visualization, and parallel volume rendering.

D. Leevers (PI) 1995-1998 European Commission $1.57M Collaborative Integrated Communications for Construction (Augmented Reality Component). This project developed augmented reality technologies for use in the construction industry. Role: co-PI ALA Application Template 01/26/2004 Page ____ Biographical Sketch