Image Processing Group Profile

Transcription

1 Profile

2 2 Published by: Faculty of Electrical Engineering and Computing University of Zagreb, Croatia Publication date: May 2004 Copyright by

3 3 Contents Contents 3 About the 5 A Short Biography of Prof. Sven Loncaric 6 Members 7 Faculty 7 Graduate Assistants 7 Graduate Students 7 Teaching Activities 8 Graduate Theses 9 Undergraduate Theses 9 Research Activities 11 National Research Partners 12 International Research Partners 12 Research and Development Projects 13 3-D Quantitative Analysis of Intracerebral Brain Hemorrhage 14 Virtual Bronchoscopy 15 WWW-based Virtual Endoscopy 16 Quantitative Analysis of Abdominal Aortic Aneurysm 17 Brain Perfusion Analysis 18 Virtual Reality Markup Language Reference Center 19 Virtual Reality Reference Center 20 Medical Image Segmentation Techniques 21 Content-Based Image Retrieval 22 Cardiac Image Analysis 23 A Web-Based Medical Teleconsultation System 24 Volume Visualization Software Modules 25 Quantitative Analysis of SPECT Brain Images 26 Image Management Techniques 27 Intelligent Systems For Medical Image Analysis 28 Deformable Models For Medical Image Analysis 29 Neural Networks For Medical Image Analysis 30 Energy Minimization Techniques For Medical Image Analysis 31 Image Registration Methods 32

4 4 Shape Description Methods 33 Face Recognition Techniques 34 SVD-Based Techniques For Image Restoration 35 Web-Based Electronic Journal Publishing Support System 36 Neural Networks For Financial Market Analysis 37 CHAIRMAN Conference Management System 38 COST B21 Project: Physiological Modelling of MR Image Formation 39 Biomedical Perfusion Image Sequence Analysis 40 Serial Registration of Pre- and Intra-operative MR Brain Images 41 Neural Networks for Ion Chromatography Optimization 42 Professional Activities 43 Conference and Workshop Organization 43 Invited Lectures 43 Journal Editorships 44 Publications 45 Journal Papers 45 Books and Book Chapters 45 Refereed Conference Papers, Refereed Conference Papers, Refereed Conference Papers, Refereed Conference Papers, Refereed Conference Papers, Refereed Conference Papers, Refereed Conference Papers, Refereed Conference Papers, 1996 and older 50

5 5 About the (IPG) at the Department of Electronic Systems and Information Processing (ESIP) was founded in 1994 by Dr. Sven Loncaric. ESIP is one of the twelve departments which constitute the Faculty of Electrical Engineering and Computing at University of Zagreb. Faculty of Electrical Engineering and Computing is one of the most prestigious scientific institutions in Croatia. University of Zagreb has been founded in 1669 and has about 50,000 undergraduate and graduate students. Faculty of Electrical Engineering and Computing has about 100 permanent faculty members in all areas of electrical and computer engineering and computer science. Information about the IPG is available on the IPG web site: For further information about the IPG activities please contact: Professor Sven Loncaric Faculty of Electrical Engineering and Computing University of Zagreb Unska Zagreb Croatia Tel: , (administrative assistant) Fax: Dr. Loncaric can also be reached at his address: sven.loncaric@fer.hr

6 6 A Short Biography of Prof. Sven Loncaric Prof. Sven Loncaric received his Diploma of Engineering degree and his M.S. degree in Electrical Engineering from University of Zagreb, Croatia, in 1985 and From , he was a Fulbright Fellow and research assistant at University of Cincinnati, USA. Prof. Loncaric received his Ph.D. degree in electrical engineering from University of Cincinnati, in From 1994 to 2001 he was a higher assistant and assistant professor at Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia where he taught undergraduate and graduate courses on Circuit Theory, System and Signal Theory, Digital Signal Processing, Digital Image Processing, Digital Image Analysis, Neural Networks, Stochastic Systems, and Virtual Reality. In 1994, he founded the at the. Since then, he has been director of the and principal investigator on several research projects and has published more than 100 publications in the area of image processing and pattern recognition with applications biomedicine and other areas. He has served as academic advisor to undergraduate and graduate students. In 2001, Prof. Loncaric joined New Jersey Institute of Technology, Department of ECE, where he was an assistant professor and Director of the Visual Information Processing Laboratory. At NJIT he taught Digital Image Processing, Microprocessors, and Microprocessor Laboratory courses. Since 2003, Prof. Loncaric has been an associate professor at Faculty of Electrical Engineering and Computing, University of Zagreb. Prof. Loncaric has served as Editor-in-Chief of an international research journal titled "Journal of Computing and Information Technology". He was Guest Editor of the Special Issue on Biomedical Image Processing and Analysis published in the Journal of Computing and Information Technology. Prof. Loncaric was founder of the series of international scientific symposia on Image and Signal Processing and Analysis (ISPA). He organized the "8th International Summer School on Image Processing" held in 2000, in Zagreb, Croatia. His research interests are in the area of digital image processing and computer vision, medical image analysis, pattern recognition, medical imaging, 3-D visualization, and virtual reality.

7 7 Members Research and development activities of the group have been carried out by a number of former and current graduate and undergraduate students. In his role as academic advisor, Dr. Loncaric has had opportunity to work with many excellent students, some of which are shown below. For a list of former undergraduate and graduate students please see IPG web site at: Faculty Sven Loncaric, director Graduate Assistants Hrvoje Bogunovic, research and teaching assistant Tomislav Petkovic, research and teaching assistant Marko Subasic, research and teaching assistant Graduate Students Hrvoje Regelja, graduate student Ivan Stajduhar, graduate student Adrijan Bozinovski, graduate student Sasa Galic, graduate student

8 8 Teaching Activities Prof. Sven Loncaric teaches undergraduate and graduate courses at Faculty of Electrical Engineering and Computing (FEEC), University of Zagreb. He is a member of the Networks, Signals, and Systems teaching group at FEEC. The group has seven teachers who teach courses in the area of signal and system theory. Prof. Loncaric has taught the following undergraduate courses: Systems and Signals Linear Networks Digital Signal Processing Digital Image Processing Neural Networks Stochastic Processes in System Theory The graduate courses currently taught by prof. Loncaric at FEEC are: Selected Topics in Image Processing Digital Image Analysis Neural Networks At New Jersey Institute of Technology, USA, Dr. Loncaric taught "Digital Image Processing", "Microprocessors", and "Microprocessor Laboratory" courses. At Faculty of Mathematics, University of Zagreb, Prof. Loncaric offers a graduate course titled "Mathematical Methods in Image Processing". As part of international teaching activities Dr. Loncaric has been junior lecturer at "3 rd IEEE EMBS Int'l Summer School on Biomedical Imaging", Berder, France, He has been invited lecturer for several years at "Int'l Summer School on Image Processing" held in Hungary and Croatia. Lecture notes and other information for all courses taught is available at the IPG web site: Graduate assistants working on their graduate theses under prof. Loncaric's mentorship participate in teaching activities as teaching assistants for Digital Image Processing Laboratory and Stochastic Systems Laboratory. Teaching assistants also give tutoring sessions in several courses taught by prof. Loncaric. The list of graduate and undergraduate students mentored by Prof. Loncaric and titles of their theses are given in the following text.

9 9 Graduate Theses Marko Subasic, "Geometric Deformable Models for Medical Image Analysis", 2003, MS Thesis Srisheela Devabhaktuni, "Elastic Brain Image Registration using Mutual Information", 2003, MS Thesis Domagoj Kovacevic, "Deformable Models for Medical Image Analysis", 2002, MS Thesis Miroslav Koncar, "Distributed Information Systems for Electronic Health Record Management", 2002, MS Thesis Ivan Vrancic, "Pattern Recognition for Financial Data Analysis", 2002, MS Thesis Zoran Biuk, "Face Recognition from Multiple Views", 2001, MS Thesis Zdravko Liposcak, "Face Recognition from Image Sequences", 2000, MS Thesis Zeljko Devcic, "System identification methods for image degradation modeling", 1998, MS Thesis Dubravko Cosic, "CT Head Image Analysis", 1997, MS Thesis Zoran Dukic, "Shape Analysis using Morphological Signature Transform", 1996, MS Thesis Undergraduate Theses Tihomir Konosić, "Graphical User Interface for Volume Registration and Visualization", 2003 Marko Dvecko, "Volume Visualization Algorithms", 2003 Andrej Dolmac, "Volume Visualization Algorithms", 2003 Tomislav Petkovic, "Methods for Visual Quality Control of Plastic Products", 2002 Marijana Grgek, "Quantitative Analysis of 3-D MR Brain Images for Volume Measurements", 2001 Ivan Stojic, "Spatio-temporal Segmentation using Simulated Annealing", 2001 Edgar Pek, "Telemedical System for Internet-based Consultations", 2001 Josip Krapac, "Computer Vision for Automated Quality Control", 2001 Hrvoje Bogunovic, "Segmentation from Image Sequences", 2001 Zeljko Tandaric, "Content-based Image Retrieval", 2001 Sinisa Skrinjar, "System for Automatic Generation of WML Pages", 2001 Ivan Ceskovic, "Quantitative Analysis of SPECT Images", 2001 Milan Matesin, "Intelligent System for Quantitative analysis of CT Images", 2001 Berislav Crkvenac, "Content-based Image Retrieval", 2000 Dejan Marcetic, "Program for 3-D image visualization", 2000

10 Darije Ramljak, "3-D volume rendering visualization methods", 2000 Tomislav Markovinovic, "2-D and 3-D object thinning methods", 1999 Marko Subasic, "3-D stereo visualization using VTK", 1999 Tomislav Petrovic, "Neural networks for image analysis", 1999 Gordana Magdic, "ANSI Z39.50 standard for information servers", 1999 Zoran Majcenic, "Neural networks for medical image segmentation", 1999 Sasa Galic, "Optical flow techniques for cardiac imaging", 1999 Miroslav Koncar, "DICOM standard for radiological PACS", 1999 Boris Svilicic, "User interface for radiological PACS", 1999 Kresimir Prpa, "Image reconstruction from projections", 1999 Ivica Borscak, "Fractal image compression", 1999 Ivan Jakab, "Digital document processing and archiving", 1998 Damir Boraska, "Web-based electronic journal publishing system", 1998 Sasa Gavrilovic, "Web-based bibliographical information system", 1998 Jasmina Delic, "Databases for library information system", 1998 Kristina Peric, "A user interface for bibliographical information system", 1998 Ozren Rajkovic, "Marching cube algorithm for volume visualization", 1997 Ivan Vrancic, "Software system for network access to library information databases", 1997 Domagoj Kovacevic, "Neural network segmentation of CT head images", 1996 Damir Cosic, "Volume vizualization methods", 1996 Damir Lisica, "Implementation of distance transform algorithms in Khoros environment", 1995 Goran Dokmanovic, "Multimedia software for geographical information systems", 1995 Kresimir Simunic, "Genetic algorithm-based image registration", 1995 Djordje Vukdragovic, "Energy minimization approaches for image segmentation", 1995 Goran Polonji, "An implementation of Hough transform algorithm", 1995 Miro Zdilar, "Percepton neural network for pattern recognition",

11 11 Research Activities brings together general expertise in the area of electronic systems and computer science and engineering. The primary emphasize in research is given to image processing, pattern recognition, computer vision, and medical imaging. Other areas of interest include 3-D visualization using virtual reality, and volume rendering methods. Computer vision and image processing have applications in many areas of human activity such as industry and medicine. The common task in all applications is a need for an image processing, e.g. to remove noise or restore image, followed by an image analysis in order to recognize certain parts of the scene and then perform image-based measurements or visualization of desired structures. Research areas in image processing and computer vision include techniques for image filtering, image enhancement, image restoration, image segmentation, image labeling, image understanding. A number of techniques is used in building the solutions for various computer vision problems including neural networks, genetic algorithms, expert systems, physics-based deformable models, clustering algorithms, optimization algorithms, and mathematical morphology. Biomedical image processing and analysis has been one of the main IPG research topics. Modern medical imaging modalities provide medical doctors with huge amount of information that requires quantitative analysis for diagnostic and interventional medical purposes. The image data can be twodimensional (2-D) or three-dimensional (3-D) and can be a static image or an image sequence. 2-D or 3-D image data is typically obtained by means of a magnetic resonance (MR) or a computed tomography (CT) scanner, or by an ultrasound scanner. The structure of interest has to be extracted using pattern recognition methods. The obtained structure can be measured and visualized for use by a medical doctor. Modern medical imaging modalities use DICOM standard to enable digital exchange of image data. Image management techniques are used to store, retrieve, and display image data and integrate this information into radiological and other hospital information systems.

12 12 National Research Partners Croatian Ministry of Science and Technology Croatian Academic and Research Network Department of Radiology, Sestre Milosrdnice University Hospital, University of Zagreb, Croatia Institute for Brain Research, Faculty of Medicine, University of Zagreb, Croatia International Research Partners Department of Neurosurgery, University of Medicine and Dentistry of New Jersey, USA Computer Vision Laboratory, University of Ljubljana, Slovenia Department of Electrical and Computer Engineering, New Jersey Institute of Technology, USA Department of Electrical and Computer Engineering, University of Iowa, USA Department of Applied Informatics, University of Szeged, Hungary Department of Radiology, University Hospital Graz, Austria

13 13 Research and Development Projects IPG topics of interest include all aspects of image processing and analysis. This includes theoretical basis and practical applications. Diverse areas such as neural networks, genetic algorithms, intelligent systems, mathematical morphology, and other are investigated and applied to problems in image processing and computer vision. A special research focus has been in medical image processing and analysis. A variety of contemporary medical imaging instruments provide physicians with a wealth of two-dimensional (2-D) and three-dimensional (3-D) image data. Interpretation of this data in everyday clinical practice is very important for medical diagnostics and intervention. Image interpretation represents a challenging task due to a large amount of acquired data. Intelligent medical image analysis is required for such quantitative image measurements. Volume visualization has played an important role in 3-D medical image processing and analysis. Volume rendering has gained a lot of popularity in recent years because of a realistic visualization of 3-D data. Virtual reality techniques and stereo visualization has enabled interactive fly-through visualizations of various body structures and organs and have been used both in education and diagnostics. In the following material, research activities of the IPG are presented with basic information about each research project. At the end of this report, the list of publications produced by IPG researchers has been provided.

14 14 3-D Quantitative Analysis of Intracerebral Brain Hemorrhage Ministry of Science and Technology, Croatia National Institutes of Health, USA Duration: May 1995-May 2001 People: Sven Loncaric, University of Zagreb, Croatia, principal co-investigator Atam P. Dhawan, University of Cincinnati, USA, principal coinvestigator Dubravko Cosic, University of Zagreb, graduate student Domagoj Kovacevic, University of Zagreb, graduate student Computed tomography (CT) allows three-dimensional (3-D) anatomical imaging of brain abnormalities such as human spontaneous intracerebral brain hemorrhage (ICH). With computerized image analysis, it is feasible to characterize the pathology of a selected volume of interest. The proposed research focuses on 3-D quantitative analysis to study the early evolution of the ICH. The underlying hypothesis is that the ICH volume and structure is related to the mortality and morbidity. Patients having ICH are scanned four times: within three hours after first symptoms, one hour later, eight hours later, and within twenty hours after first symptoms. During the course of the illness, 3-D changes in ICH volume and structure can be observed and analyzed. The important ICH features are volume, position in space, and shape of primary and edema region. Preliminary studies indicate that the ICH volume is significant for the survival of the patient. The position in space must be measured with respect to an invariant 3-D coordinate system so that the movement of the ICH across the scans can be determined. We have developed 3-D spatially weighted region growing algorithms with adaptive clustering for segmentation of ICH regions. Publications: D. Cosic and S. Loncaric. Computer System for Quantitative Analysis of ICH from CT Head Images. Proceedings of the 19th Annual International Conference of IEEE Engineering in Medicine and Biology Society, Chicago, USA, S. Loncaric and D. Kovacevic. A Method for Segmentation of CT Head Images. Proceedings of the 9th International Conference on Image Analysis and Processing, pp , Florence, Italy, D. Cosic and S. Loncaric. A Rule-based Labeling of CT Head Image. Proceedings of the 6th Conference on Artificial Intelligence in Medicine Europe, Grenoble, France, Lecture Notes in Artificial Intelligence, pp , Springer-Verlag, S. Loncaric and D. Cosic and A. P. Dhawan. Hierarchical Segmentation of CT Head Images. Conference Proceedings of the 18th Annual International Conference of IEEE Engineering in Medicine and Biology Society, Amsterdam, Netherlands, S. Loncaric and D. Cosic and A. P. Dhawan. Segmentation of CT Head Images. Computer Assisted Radiology '96 Proc. of the International Symposium on Computer and Communication Systems for Image Guided Diagnosis and Therapy, pp , Paris, France, 1996.

15 15 Virtual Bronchoscopy Central European Exchange Program for University Studies (CEEPUS) Duration: September 1997 August 1998 People: Sven Loncaric, University of Zagreb, Croatia, principal co-investigator Erich Sorantin, Karl-Franzens University of Graz, Austria, principal co-investigator Attila Kuba, University of Szeged, Hungary, principal co-investigator Tomislav Markovinovic, University of Zagreb, Croatia, undergraduate student Tomislav Petrovic, University of Zagreb, Croatia, undergraduate student Darije Ramljak, University of Zagreb, Croatia, undergraduate student Virtual bronchoscopy is a visualization technique based on computed tomography (CT) or magnetic resonance (MR) scanning of the organ of interest and extracting its 3-D model. The 3-D model can be interactively examined or fly-through visualizations along a path through the structure can be generated by volume or surface rendering. Such an approach is called virtual endoscopy and it has several advantages over conventional endoscopy where an instrument is inserted through a natural or minimally invasive opening on the human body. The advantages of virtual endoscopy are non-invasiveness of the technique and ability to move through structures where physical endoscope cannot enter. The main disadvantage of virtual endoscopy is that the color and the texture of organ surface cannot be restored from scanned images. The aim of the project is to realize the software application for virtual bronchoscopy i.e. for virtual endoscopy of the bronchial tubes. The input images are acquired by CT scanning. Image analysis must be conducted to determine the location of the bronchial tubes and to extract them. The goal of the procedure is to create a fly-through animation along the medial axis (skeleton) of the bronchial tubes. The 3-D skeleton is computed from segmented volume. The animation is created by volume rendering along the computed medial axis. Publications: S. Loncaric and T. Markovinovic and T. Petrovic and D. Ramljak and E. Sorantin. Construction of Virtual Environment for Endoscopy. Proceedings of the IEEE International Conference on Multimedia Computing and Systems, Florence, Italy, S. Loncaric and D. Ramljak and T. Markovinovic and E. Sorantin. VRMLbased Model Creation for Virtual Bronchoscopy. First Croatian Symposium on Computer Assisted Surgery, Zagreb, Croatia, S. Loncaric. Virtual Reality in Medicine: An Overview. First Croatian Symposium on Computer Assisted Surgery, Zagreb, Croatia, 1999.

16 16 WWW-based Virtual Endoscopy Central European Exchange Program for University Studies (CEEPUS) Duration: September 1998 August 1999 People: Sven Loncaric, University of Zagreb, Croatia, principal co-investigator Erich Sorantin, Karl-Franzens University of Graz, Austria, principal co-investigator Attila Kuba, University of Szeged, Hungary, principal co-investigator Tomislav Markovinovic, University of Zagreb, Croatia, undergraduate student A WWW-based implementation has been realized where the user interacts with the system by means of WWW browser. Such an approach provides platform independence. Segmentation and visualization is computed on-the-fly on the WWW server side. The results are transferred to the WWW client and presented to the user in the form of a MPEG movie (fly-through animation) or in the form of a VRML model which can be examined in a VRML-enabled WWW browser. The animation is produced by volume rendering of the volume data along the fly-through path. A speed-ptimized volume rendering algorithm has been used to enable higher computation speed. The fly-through path is computed automatically from the segmented bronchial structure using a distance transformbased algorithm for bronchial skeleton computation. User assistance is required for selection of the desired main branch of the skeleton to be used for fly-through path. The computed skeleton is smoothed using a B-spline interpolation technique to provide a smooth animation. The software has been realized on a SUN workstation. Publications: S. Loncaric, T. Markovinovic, WWW-based virtual endoscopy, submitted to Medical Informatics Europe, Hannover, Germany, 2000

17 17 Quantitative Analysis of Abdominal Aortic Aneurysm Central European Exchange Program for University Studies (CEEPUS) Duration: September 1999 August 2000 People: Sven Loncaric, University of Zagreb, Croatia, principal co-investigator Erich Sorantin, Karl-Franzens University of Graz, Austria, principal co-investigator Attila Kuba, University of Szeged, Hungary, principal co-investigator Domagoj Kovacevic, University of Zagreb, graduate student Marko Subasic, University of Zagreb, graduate student Abdominal aortic aneurysm (AAA) is a vascular disease which can be treated by open surgical procedure or by a minimally invasive surgical procedure. In order to correctly select the size and shape of the stent graft device it is necessary to perform accurate measurements of the aneurysm. This can be done by CT or MR scanning. In this project a spiral CT scanner has been used for acquiring the chest images containing the aneurysm. Deformable models have been used for 3-D segmentation of AAA. In the first approach, a classical snake algorithm has been used in 2-D slices with additional forces acting between the slices. In the second approach, a level-set algorithm, according to Sethian, has been used for implementation of deformable model and its evaluation. The software has been implemented in MATLAB and C languages. Publications: S. Loncaric and D. Kovacevic and E. Sorantin. Semi-automatic Active Contour Approach to Segmentation of Computed Tomography Volumes. Proceedings of SPIE Medical Imaging, San Diego, USA, 2000 M. Subasic and S. Loncaric. 3-D deformable model for aneurysm segmentation, submitted to Medical Informatics Europe conference, Hannover, Germany, 2000 D. Kovacevic and S. Loncaric and E. Sorantin. Deformable Models for Medical Image Segmentation. First Croatian Symposium on Computer Assisted Surgery, Zagreb, Croatia, 1999.

18 18 Brain Perfusion Analysis Central European Exchange Program for University Studies (CEEPUS) Duration: September 2000 August 2001 People: Sven Loncaric, University of Zagreb, Croatia, principal co-investigator Nada Besenski, University of Zagreb, Croatia, principal co-investigator Erich Sorantin, Karl-Franzens University of Graz, Austria, principal co-investigator Attila Kuba, University of Szeged, Hungary, principal co-investigator The project is focused on quantitative image-based measurement of diffusion and perfusion in brain ischemy. The brain images are acquired using magnetic resonance (MR) scanning. The analysis must be performed on single images as well as on image sequences to track time changes in patients with degenerative brain diseases. It is very important for diagnostic, therapeutic, and prospective reasons to determine the size and age of ischemia, and the size of penumbra in ischemic lesion. New MR imaging techniques such as diffusion weighted imaging and echoplanar perfusion enable such measurements. Image processing and analysis methods will be based on neural networks, expert systems, mathematical morphology, and other techniques. A clinical study will be conducted to determine the validity of the method.

19 19 Virtual Reality Markup Language Reference Center Croatian Academic and Research Network (CARNet) Duration: January 1996 December 1997 People: Sven Loncaric, University of Zagreb, Croatia, principal investigator Tomislav Markovinovic, University of Zagreb, Croatia, undergraduate student Tomislav Petrovic, University of Zagreb, Croatia, undergraduate student Darije Ramljak, University of Zagreb, Croatia, undergraduate student Berislav Crkvenac, University of Zagreb, Croatia, undergraduate student Recently, virtual reality has become an exciting research area with many applications. Virtual Reality Modeling Language (VRML) is a quickly developing tool for description of 3-D objects which combines virtual reality technologies with network technologies. The aim of this project is to establish a referral center for VRML technology in Croatia. The project duration is one-year, starting in February 1997, and ending in February The project goals are: to establish a referent educational center for VRML technology to test hardware and software for virtual reality to test VRML modelers and browsers on PC (Windows 95, Windows NT, and Linux) and Sun Solaris platforms to develop a model of Croatian territory in VRML format Publications: S. Loncaric and T. Markovinovic and T. Petrovic and D. Ramljak and E. Sorantin. Construction of Virtual Environment for Endoscopy. Proceedings of the IEEE International Conference on Multimedia Computing and Systems, Florence, Italy, S. Loncaric and D. Ramljak and T. Markovinovic and E. Sorantin. VRMLbased Model Creation for Virtual Bronchoscopy. First Croatian Symposium on Computer Assisted Surgery, Zagreb, Croatia, S. Loncaric. Virtual Reality in Medicine: An Overview. First Croatian Symposium on Computer Assisted Surgery, Zagreb, Croatia, 1999.

20 20 Virtual Reality Reference Center People: Croatian Academic and Research Network (CARNet) 1998 Sven Loncaric, University of Zagreb, Croatia, principal investigator Marko Subasic, University of Zagreb, Croatia, graduate student The aim of the project is to establish a center for virtual reality (VR) technology where users from academia and industry will have opportunity to learn about the VR technology and be able to employ these techniques in various fields of human activity. The VR center will include a fast graphics workstation with projector and a rear projection screen, stereo glasses, and a tracking device. This VR setup will be used for various VR visualization applications in mechenical engineering, chemical engineering, art, and other areas. Researchers from other academic institutions will have opportunity to bring their VRML models to the VR workstation and examine them.

21 21 Medical Image Segmentation Techniques Ministry of Science and Technology, Croatia Ministry of Science and Technology, Slovenia Duration: January 1996 December 1997 People: Sven Loncaric, University of Zagreb, Croatia, principal investigator Franc Solina, University of Ljubljana, Slovenia, principal co-investigator The role of the project is to encourage collaboration between the two research groups, in the area of computer vision and image processing, Computer Vision Laboratory at University of Ljubljana and at University of Zagreb. In particular the aim is to propose new medical image segmentation techniques based on neural networks, optimization algorithms, clustering techniques, expert systems, and deformable models.

22 22 Content-Based Image Retrieval People: Pilot project Sven Loncaric, University of Zagreb, Croatia, coordinator Berislav Crkvenac, University of Zagreb, Croatia, undergraduate student Zeljko Tandaric, University of Zagreb, Croatia, undergraduate student Content-based image retrieval has become very important, particularly in recent years with the growth of Internet and a huge number of images on the web sites. This fact demonstrates the need for tools for searching the Internet for images by content.tools for indexing and retrieving the textual information from the Internet have been developed but development of content-based image retrieval have not been well researched. Tools for content-based image retrieval typically use color, shape, and texture features to classify images and define criteria of similarity. The difficulty is in defining a query image. In this project, we propose to develop a neural network based method for training the system to learn the similarity criteria for each particular application. Once the neural network is trained it is used for content-based search by comparing each image from the image database with query (example) image. The method will be tested on a class of medical images.

23 23 Cardiac Image Analysis People: Pilot project Sven Loncaric, University of Zagreb, Croatia, principal co-investigator James Duncan, Yale University, USA, principal co-investigator Sasa Galic, University of Zagreb, Croatia, graduate student Tvrtko Macan, University of Zagreb, Croatia, graduate student Zoran Majcenic, University of Zagreb, Croatia, graduate student Marko Subasic, University of Zagreb, Croatia, graduate student Heart diseases such as heart ischemia change the heart motion because necrosed regions of the heart have different viscoelastic properties. To distinguish between the normal and the ischemic heart and determine the nature of pathology it is necessary to perform estimation of the left ventricle (LV) motion. Once LV motion is known it is possible to perform strain analysis using heart tissue modelling and finite-element method for numerical solution of the obtained system of diferential equations. In this project we investigate new techniques for LV segmentation and motion estimation. We pay special attention to spatio-temporal methods for 2-D and 3-D segmentation and motion estimation utilizing deformable models, neural networks, optical flow techniques, correlation techniques, and simulated annealing. Imaging modalities that are used are MR ECG-gated scanning and ultrasound. Publications: S. Loncaric and T. Macan. Point-constrained optical flow for LV motion detection. Proceedings of SPIE Medical Imaging, San Diego, USA, 2000, to be presented. Z. Majcenic and S. Loncaric. An algorithm for spatio-temporal heart segmentation. Proceedings of SPIE Medical Imaging, San Diego, USA, 2000, to be presented.

24 24 A Web-Based Medical Teleconsultation System People: Publications: Ministry of Science and Technology (An IT project) Sven Loncaric, University of Zagreb, Croatia, principal co-investigator Anamarija Margan, Cres, Croatia principal co-investigator Miroslav Silovic, University of Zagreb graduate student Edgar Pek, University of Zagreb undergraduate student Medical consultations between ambulances in rural areas and large medical centers are necessary for high quality health care. In particular, medical consultations between Croatian islands of Cres and Losinj and mainland medical centers in Rijeka and Zagreb are necessary for several reasons. Often, patient transport is not acceptable due to a high risk for the patient. In addition, transport generates additional expense and is often impossible in winter months due to weather and sea conditions, not to mention discomfort to the patient. In this paper, we describe a concept of a web-based system for medical teleconsultation. The system will be applied to communication between the islands of Cres, Losinj, and neighboring islands and university hospitals in Zagreb and Rijeka. The system consists of a teleconsultation module and a electronic Internet-based health record module. User interface is implemented in HTML and Phyton languages. WWW access to the system ensures platform independence, i.e. the system can be accessed using an ordinary Internet browser from any computer platform. The system is currently under development. S. Loncaric and E. Pek and M. Silovic and A. Margan. Virtual Polyclinic: A Web-based system for medical teleconsultation. Proceedings of Emerging Technologies and Life Sciences: Medicine and Communication (MEDICOM), Remagen, Germany, S. Loncaric and M. Silovic and E. Pek and A. Margan. A Web-based system for medical teleconsultation: First results. Proceedings of Telemedicine and Health Informatics, Milano, Italy, 2000.

25 25 Volume Visualization Software Modules Ministry of Science and Technology, Croatia VAMS Duration: October People: Sven Loncaric, University of Zagreb, Croatia, coordinator Josip Krapac, University of Zagreb, Croatia, undergraduate student Dejan Marcetic, University of Zagreb, Croatia, undergraduate student Darije Ramljak, University of Zagreb, Croatia, graduate student The goal of this project is to develop software modules in the form of dynamic libraries (DLLs) for Windows environment for use from a teleradiological application. The purpose of these modules is visualization of 3-D data such as computed tomography (CT) and magnetic resonance (MR) scans. The software modules must enable several visualization types. The first and the simpliest type is multi-planar visualization showing to the radiologist axial, coronal, and sagital views of the volume. The user must be able to navigate in all three planes and interpolation must be implemented to increase the resolution in Z direction. The second type of visualization is using a VRML surface model of the organ of interest. This type of visualization enables interactive examination of the volume of interest by means of a VRML viewer. The user can navigate through the volume in real time. The third type of visualization uses volume rendering for the best visualization of volume. The price to be paid for this are increased computational requirements and resulting lower speed. The user can specify a 3-D path for fly-through animation and the module will compute the rendered frames along the path and create a MPEG movie. The software modules will be integrated with the main program and tested by users. Graphical user interfaces for each module will be developed to enable effective communication with the software with minimal amount of interaction. Publications: D. Cosic and S. Loncaric and O. Rajkovic. A Method for 3-D Image Representation Using a Surface Model. Proceedings of the IEEE 9th Mediterranean Electrotechnical Conference, pp , Tel-Aviv, Israel, 1998.

26 26 Quantitative Analysis of SPECT Brain Images pilot project Duration: October People: Sven Loncaric, University of Zagreb, Croatia, coordinator Srecko Loncaric, Faculty of Medicine, University Hospital Rebro, coordinator Ivan Ceskovic, University of Zagreb, Croatia, undergraduate student Single Photon Emission Computed Tomography (SPECT) medical imaging modality has been used for functional imaging of various organs in the human body including brain. Functional imaging provides information about the function of an organ as opposed to modalities that provide anatomical information about the structures such as computed tomography (CT). In emision tomography a radioactive contrast medium is injected into the patient and radiation that is emitted as a result of radioactive decay is detected by a camera. A wide spectrum of different contrast media exists that are used for different organs and that reveal various aspects of the organ function. In SPECT, the 3-D organ is reconstructed from multiple projections to provide spatial information about the intensity of activity in different parts of the organ. SPECT imaging is used in studies of tumors, thyroid glands, and other organs. A disadvantage of SPECT with respect to CT is reduced resolution what makes image processing procedures more difficult. The goal of this project is to develop image processing and analysis algorithms and procedures for measurements of tumor size. The procedures must enable temporal studies to track tumor changes in time and evaluate the effect of therapy to the patient. The acquired images are obtained from the SPECT scanner in digital form by means of DICOM standard. The program will be developed for Windows PC environment and for UNIX environment.

27 27 Image Management Techniques pilot project Duration: October People: Sven Loncaric, University of Zagreb, Croatia, coordinator Miroslav Koncar, University of Zagreb, Croatia, graduate student Dzumridin Ibraimi, University of Zagreb, undergraduate student Image management techniques play an important role in radiological information systems and many other non-medical applications. These techniques are basis for image archiving, transfer, and retrieval of images from image databases in a networked environment. The goal of this project is to develop an experimental picture archiving and communication system (PACS) for DICOM images. The WWW user interface is used. In this way, any graphical workstation equipped a quality monitor and with a WWW browser can serve as a radiological viewing station. The system can be accessed anywhere from the hospital Intranet. Experimental image archive will be based on CTN DICOM server implementation. DICOM import and export tools in MATLAB environment will be developed to be used by MATLAB image processing and analysis software modules. DICOM server WWW server WWW client LAN CT imaging modalities MR WWW client Publications: M. Koncar and S. Loncaric. WWW-based radiology image management system. to be presented at Medical Informatics Europe conference, Hannover, Germany, 2000 M. Koncar and S. Loncaric. WWW-based radiological PACS. Proceedings of the Croatian Medical Informatics Symposium, pp , Zagreb, Croatia, 1999.

28 28 Intelligent Systems For Medical Image Analysis pilot project Duration: October People: Sven Loncaric, University of Zagreb, Croatia, coordinator Domagoj Kovacevic, University of Zagreb, Croatia, graduate student Milan Matesin, University of Zagreb, Croatia, undergraduate student Image interpretation is a challenging task. The reason for this is that for accurate image interpretation it is necessary segment the scene into its parts and to perform recognition of objects in the scene. This task requires intelligence to be successfully performed. Implementation of intelligent procedures on a computer constitutes the field of artificial intelligence. In the area of artificial intelligence, a number of different methods have been studied. In particular expert systems have been used in many image understanding problems. A subgroup of expert systems are rule-based expert systems which are a useful tool for solution of specialized image understanding problems. In this project we will study intelligent systems that are useful for implementation of artificial image interpretation systems. After the initial image processing, the image is segmented into some uniform regions. A set of features is assigned to each region describing the properties of the region and its relation to the neighboring regions. Rules are formed to describe the knowledge about the structure of the image, i.e. what are the possible and meaningful configurations of regions. The task of rule-based expert system is to use input regions and rules to perform image labeling, i.e. to identify which region represents what part of the scene. Nonfuzzy and fuzzy rule-based expert systems will be investigated. The goal of the project is to develop a software for intelligent interpretation of medical images of particular class such as brain images, abdominal images, etc. The software will utilize the concepts of intelligent systems including rule-based expert systems and other techniques. Publications: S. Loncaric and D. Kovacevic and D. Cosic. Fuzzy Expert System for Edema Segmentation. Proceedings of the IEEE 9th Mediterranean Electrotechnical Conference, pp , Tel-Aviv, Israel, D. Cosic and S. Loncaric. A Rule-based Labeling of CT Head Image. Proceedings of the 6th Conference on Artificial Intelligence in Medicine Europe, Grenoble, France, Lecture Notes in Artificial Intelligence, pp , Springer-Verlag, 1997.

29 29 Deformable Models For Medical Image Analysis Central European Exchange Program for University Studies (CEEPUS) Duration: October 1998-October 2001 People: Sven Loncaric, University of Zagreb, Croatia, coordinator Erich Sorantin, Karl-Franzens University Graz, Austria, coordinator Marijana Grgek, University of Zagreb, Croatia, undergraduate student Domagoj Kovacevic, University of Zagreb, Croatia, undergraduate student Marko Subasic, University of Zagreb, Croatia, graduate student Physics-based deformable models are a powerful tool for segmentation of 2-D and 3-D structures in medical image analysis. In this project we investigate medical image analysis techniques based on deformable models. In deformable models, a model of shape border or surface is deformed under the influence of external and internal forces acting on the model. The deformation process is governed by the motion partial differential equations which have to be solved in orer to determine the dynamics of the model. In image analysis we are only interested in the final solution and not in the intermediate time steps. We investigate different implementations of the algorithm such as classical snake and baloon based approach and alternative level-set implementation of the algorithm. The goal of the project is to develop algorithms and software modules for specific problems in medical image analysis including segmentation of the brain ventricle, abdominal aortic aneurysm segmentation, and heart left ventricle segmentation. Publications: M. Subasic and S. Loncaric. 3-D deformable model for aneurysm segmentation. submitted to Medical Informatics Europe conference, Hannover, Germany, 2000 S. Loncaric and D. Kovacevic and E. Sorantin. Semi-automatic Active Contour Approach to Segmentation of Computed Tomography Volumes. Proceedings of SPIE Medical Imaging, San Diego, USA, 2000, to be presented. D. Kovacevic and S. Loncaric. Deformable Contour-Based Method For Medical Image Segmentation. Proceedings of the 21st Int'l Conference Information Technology Interfaces, pp , Pula, Croatia, D. Kovacevic and S. Loncaric and E. Sorantin. Deformable Models for Medical Image Segmentation. First Croatian Symposium on Computer Assisted Surgery, Zagreb, Croatia, 1999.

30 30 Neural Networks For Medical Image Analysis Ministry of Science and Technology, Croatia Duration: January People: Sven Loncaric, University of Zagreb, Croatia, coordinator Ivan Stojic, University of Zagreb, Croatia, undergraduate student Domagoj Kovacevic, University of Zagreb, Croatia, undergraduate student Segmentation of computed tomography (CT) images is an important step in many medical image analysis systems. Image segmentation is a challenging step since the region of interest is often not easily recognizable. Neural network approach results in a relatively fast and accurate automatic segmentation and has been successfully used in other medical image analysis problems. In particular, multilayer feed-forward networks using 2-D receptive fields have been used for image segmentation. In this project, we discuss the problem of receptive field shape and size selection and the corresponding image feature set selection. Several receptive fields are proposed, tested, and quantitatively compared. For a given receptive field, a number of different features is defined and compared with respect to segmentation performance. Principal component analysis has been used for selection of the most relevant features as well as transformation of image features for input into neural network. Publications: S. Loncaric and T. Petrovic. Receptive field shape and feature set selection in neural networks for image segmentation. submitted to ICPR, Barcelona, Spain, 2000 D. Kovacevic and S. Loncaric. CT Image Labeling Using Hopfield Neural Network. Proceedings of the IEEE 9th Mediterranean Electrotechnical Conference, pp , Tel-Aviv, Israel, S. Loncaric and D. Kovacevic. A Method for Segmentation of CT Head Images. Proceedings of the 9th International Conference on Image Analysis and Processing, pp , Florence, Italy, D. Kovacevic and S. Loncaric. Radial Basis Function-based Image Segmentation using a Receptive Field. Proceedings of the Tenth Annual IEEE Symposium on Computer-Based Medical Systems, pp , Maribor, Slovenia, 1997.

31 31 Energy Minimization Techniques For Medical Image Analysis Ministry of Science and Technology, Croatia Duration: October People: Publications: Sven Loncaric, University of Zagreb, Croatia, coordinator Zoran Majcenic, University of Zagreb, Croatia, undergraduate student Energy minimization techniques are a useful tool for many image processing and analysis tasks. In this approach the problem is formulated in terms of an energy function which has to be minimized. The solution to this minimization problem represents the solution of the original image analysis problem. Using such an approach image analysis problems are reduced to energy function minimization problems. Function minimization problems can be solved using a number of numerical techniques. A major problem in such algorithms is an unwanted convergence to a one of the local minimas. To avoid this problem a number of stochastic techniques have been introduced which search the optimization space in stochastic manner to avoid the premature convergence to a local minimum. The goal of the project is to develop new and computationally efficient SA algorithms for 2-D and 3-D image analysis problems such as image segmentation and motion estimation. The algorithms have been applied to the problem of brain image segmentation, and heart motion estimation and segmentation. Multiresolution SA algorithms have been developed to reduce the complexity and speed-up computations. Z. Majcenic and S. Loncaric. An algorithm for spatio-temporal heart segmentation. Proceedings of SPIE Medical Imaging, San Diego, USA, 2000, to be presented. S. Loncaric and Z. Majcenic. An efficient SA algorithm for brain hemorrhage image analysis. First Croatian Symposium on Computer Assisted Surgery, Zagreb, Croatia, S. Loncaric and Z. Majcenic. Multiresolution simulated annealing for brain image analysis. Proceedings of SPIE Medical Imaging, Vol. 3661, pp , San Diego, USA, Z. Majcenic and S. Loncaric. CT image labeling using simulated annealing algorithm. Proceedings of the IX European Signal Processing Conference, Vol. 4, pp , Island of Rhodos, Greece, S. Loncaric and Z. Majcenic. Multiresolution CT Head Image Analysis using Simulated Annealing. Proceedings of the 20th Int l Conference Information Technology Interfaces, pp , Pula, Croatia, S. Loncaric and Z. Majcenic. Multiresolution CT Head Image Analysis using Simulated Annealing. Proceedings of the 12th International Symposium Computer Assisted Radiology and Surgery, pp , Tokyo, Japan, 1998.

32 32 Image Registration Methods Ministry of Science and Technology, Croatia Duration: October People: Sven Loncaric, University of Zagreb, Croatia, coordinator Kresimir Simunic, University of Zagreb, Croatia, graduate student Image registration or matching techniques are important in many image analysis applications such as multispectral satellite imaging or multimodality medical imaging. In multi-modality medical imaging multiple images of a single object are made each of which contains distinct information about the object such as its anatomy or function. The matching problem consists in determining the unknown transform parameters required to map one image to match the other image. Most image matching research efforts have been concentrated on algorithms for matching of images containing a fullview of the object structure to be matched while partial-view algorithms have received a limited attention. There are many applications where a partial object matching is required, either because one of the objects is occluded, or because of imaging modality limitations. For instance, the positron emission tomography (PET) imaging modality gives a partial view while the magnetic resonance (MR) modality provides a complete 3-D view of the brain.the goal of the project is to investigate rigid and non-rigid, 2-D and 3-D, full-view and partial-view image registration techniques. Publications: K. Simunic and S. Loncaric. A Genetic Search-Based Partial Image Matching. Proceedings of the 2nd IEEE Int l Conference on Intelligent Processing Systems, pp , Gold Coast, Australia, S. Loncaric and A. P. Dhawan. A Near-Optimal Morphological MR-PET Brain Registration. Scientific Programme and Abstracts of the 9th Annual European Congress of Radiology, pp , Vienna, Austria, S. Loncaric and A. P. Dhawan. 3-D Brain Image Registration Using Optimal Morphological Processing and Iterative Principal Axis Transform. Proceedings of the 16th Annual International Conference of IEEE Engineering in Medicine and Biology Society, Paris, France, S. Loncaric and A. P. Dhawan. 3-D Brain Image Registration Using Optimal Morphological Processing. Proceedings of the 10th International Symposium on Biomedical Engineering, pp , Zagreb, Croatia, 1994.

33 33 Shape Description Methods Ministry of Science and Technology, Croatia Duration: October People: Sven Loncaric, University of Zagreb, Croatia, coordinator Atam P.Dhawan, University of Cincinnati, USA coordinator Zoran Djukic, University of Zagreb, Croatia, graduate student The shape of the object is a silhouette of the object (e.g. obtained by illuminating the object by an infinitely distant light source). A shape description method generates a descriptor vector (also called a feature vector) from the given shape (binary image). The required properties of a shape description scheme are invariance to translation, scale, and rotation. This is required because these three transformation do not change the shape of the object. The problem of shape description is one of the most important problems in image analysis. It has been pursued by many authors and there is a huge amount of work related to it. The goal of the project is to investigate new techniques for description of shape using various concepts such as deformable models, mathematical morphology, image transforms, and other techniques. In particular, we are interested in techniques that have perceptual basis in a human visual perception system. Publications: S. Loncaric. A Survey of Shape Analysis Techniques. Pattern Recognition, Vol. 31, No. 8, pp , S. Loncaric and A. P. Dhawan. Near-optimal MST-based Shape Description using Genetic Algorithm. Pattern Recognition, vol 28, pp , S. Loncaric. Morphological Signature Transform for Shape Representation and Matching. (Ph.D. Thesis), Univ. of Cincinnati, ECE Department, Z. Dukic and S. Loncaric. An Object-oriented Implementation of Morphological Operations using Element Vector Representation. Journal of Computing and Information Technology, vol. 4, pp , S. Loncaric. Shape Description Methods: A Review. Proceedings of the 40th Anniversary Conference KoREMA, pp , Zagreb, Croatia, Z. Dukic and S. Loncaric. An Object-Oriented Implementation of Morphological Operations Using Element Vector Representation. Proceedings of the 18th International Conference Information Technology Interfaces, pp , Pula, Croatia, Z. Dukic and S. Loncaric. Erosion and Dilation based Morphological Signature Transform for Shape Description. SPIE Electronic Imaging '97, Nonlinear Image Processing VIII, San Jose, USA, S. Loncaric and A. P. Dhawan. A morphological signature transform for shape description. Pattern Recognition, vol 26, pp , 1993.

34 34 Face Recognition Techniques pilot project Duration: January People: Sven Loncaric, University of Zagreb, Croatia, coordinator Zdravko Liposcak, University of Zagreb, Croatia, graduate student The problem of face recognition appears in many applications such as image databases for security and law-enforcement purposes. Face recognition can be performed using the binary face profiles, or the gray scale or color images of the face, or using the image sequences such as those found in the video databases. Image sequences are useful because they provide more information about the person than a single image. Recently, intensive research efforts have been put into multimedia databases and content-based retrieval and face recognition is just a special case of content based-image retrieval. The goal of the project is to investigate face recognition techniques and to pay special attention to techniques that utilize image sequences. We have investigated morphological techniques and scale-space techniques for face recognition. Eigenface techniques are investigated for use with sequences containing human faces. Software for face recognition that implements the developed algorithms will be developed. Publications: Z. Liposcak and S. Loncaric. Face recognition from profiles using morphological operations and boundary shape representation. Proceedings of the Workshop on Recognition, Analysis, and Tracking of Faces and Gestures in Real-Time Systems, Kerkyra, Greece, Z. Liposcak and S. Loncaric. A scale-space approach to face recognition from profiles. Proceedings of the 8th International Conference on Computer Analysis of Images and Patterns, Ljubljana, Slovenia, Z. Liposcak and S. Loncaric. Face Recognition from Profiles Using Morphological Signature Transform. Proceedings of the 21st Int'l Conference Information Technology Interfaces, pp , Pula, Croatia, 1999.

35 35 SVD-Based Techniques For Image Restoration Ministry of Science and Technology, Croatia Duration: October 199 People: Sven Loncaric, University of Zagreb, Croatia, coordinator Zeljko Devcic, University of Zagreb, Croatia, graduate student Image degradation manifests itself as a blurring of image and adding of noise to the image. It is caused by imperfection of imaging systems and inappropriate imaging conditions. Image blurring is modeled as a 2-D convolution of image with non-causal pointspread function (PSF). Blur identification is determination of an unknown PSF, and is the most important step towards a successful digital image restoration. Our blur identification algorithm operates on singular vectors of degraded image. We identify components of singular value decomposition (SVD) of the unknown PSF. The first (left and right) singular vectors of PSF can be estimated using corresponding first few singular vectors of degraded image. When S/N ratio is high, it is possible to estimate the second singular vectors of PSF. Proposed procedure is especially applicable at lower S/N ratios. Restoration res ults on simulated and real world images prove the validity of the approach. Publications: Z. Devcic and S. Loncaric. SVD-based image restoration technique. to be presented at ICASSP conference, Istanbul, Turkey, 2000 Z. Devcic and S. Loncaric. SVD Block Processing for Non-linear Image Noise Filtering. Journal of Computing and Information Technology, Vol. 7, No. 3, pp , Z. Devcic and S. Loncaric. Non-linear image noise filtering algorithm based on SVD block processing. Proceedings of the European Conference on Circuit Theory and Design, Vol. 2, pp , Stresa, Italy, Z. Devcic and S. Loncaric. SVD Block Processing for Non-linear Image Noise Filtering. Proceedings of the 20th Int l Conference Information Technology Interfaces, pp , Pula, Croatia, 1998.