NVPHBV - Fall Meeting 2009

NVPHBV - Fall Meeting 2009

NVPHBV Fall Meeting - November 16, 2009 Location: Leiden University Medical Center, Albinusdreef 2, 2333 ZA, Leiden. Lecture hall 4 (LUMC main building, zone K, first floor).

PROGRAM 10:15-10:45

Reception

10:45-10:50

Opening

10:50-11:10

Abstract session 1: Object recognition 3D Face Modeling for Recognition Frank B. ter Haar. TNO Defence, Security and Safety.

11:10-11:30

Object Recognition for person and car detection Rob Wijnhoven. ViNotion BV and Technical University Eindhoven.

11:30-11:50

Text spotting in unconstrained environments Jochem van Vroonhoven. Prime Vision and University Utrecht.

12:50-12:10

Real-time detection of surgical instruments in laparoscopic video: Towards automatic logging of surgical steps in the operating room Loubna Bouarfa, Maja Rudinac, Oytun Akman, Jenny Dankelman, Pieter Jonker. Dept. Bio-mechanical Engineering, Delft University of Technology.

12:10-13:10

LUNCH

Abstract session 2: Structure analysis 13:10-13:20

Computer vision and pattern recognition activities at HBO's in the Netherlands Jaap van de Loosdrecht, Kenniscentrum Computer Vision Lab, Noordelijke Hogeschool Leeuwarden.

13:20-13:40

Elementary multi-scale Riemann-Finsler geometry. Applications to tensor valued medical image analysis Laura Astola. Centre for analysis, scientific computing and applications. Technical University Eindhoven.

November 16, 2009 - Leiden University Medical Center

1


13:40-14:00

Dual tensor atlas generation based on a cohort of coregistered non-HARDI datasets Matthan W.A. Caan, C.A. Sage, M.M. van der Graaf, C.A. Grimbergen, S. Sunaert, L.J. van Vliet and F.M. Vos Dept Radiology, Academic Medical Center, University of Amsterdam, NL. Quantitative Imaging Group, Delft University of Technology, Dept. Radiology, University Hospitals of the Catholic University of Leuven, Belgium.

14:00-14:20

Tractography in anatomical images of wrist and knee: regularization of ligament orientation estimates

Martijn van de Giessen1,2, Frans M. Vos1,3, Laurens den Haan1, Kees A. Grimbergen2, Lucas J. van Vliet1, Geert J. Streekstra2. 1Delft University of Technology, Faculty of Applied Sciences, Quantitative Imaging Group Academic Medical Center, Depts. 2Biomedical Engineering and Physics, 3 Radiology.

14:20-14:50

COFFEE / TEA

14:50-15:10

Abstract session 3: Biomedical image processing DNA deformations near charged surfaces: Electron and atomic force microscopy views

F. G. A. Faas1, B. Rieger1, L. J. van Vliet1, D. I. Cherny.2 Dept. Imaging Science and Technology Delft University of Technology; 2Dept. Biochemistry, University of Leicester, Leicester, UK 1

15:10-15:30

Correspondence free 3D statistical shape model fitting to sparse X-ray projections

N. Baka1,2, W.J. Niessen2, B.L.Kaptein3, T. van Walsum2, L. Ferrarini1, J.H.C. Reiber1, B.P.F. Lelieveldt1. 1Division of Image Processing, Dept. Radiology, Leiden 2 University Medical Center (LUMC), Leiden. Biomedical Imaging Group Rotterdam, 3 Erasmus Medical Center, Rotterdam Dept. of Orthopaedic Surgery, LUMC.

15:30-15:50

Integrated visualization and analysis of coronary arteries and myocardial perfusion H.A.Kirisli, Dept. Medical Informatics & Radiology, Erasmus MC, Rotterdam.

15:50-16:10

16:10

Coronary intervention planning by fusing angiogram and IVUS

Shengxian Tu1, Gerhard Koning1, Zheng Huang2, Kai Cui2, Andrei Rares1, Jasper P. Janssen1, Johan H. C. Reiber1. 1LKEB, Dept. Radiology, Leiden University Medical Center; 2Dept. Cardiology, Nanfang Hospital, China.

Adjourn

16:15 – 17:00 Reception, Hepato Bar (Ground floor, LUMC)


2


ABSTRACTS

3D Face Modeling for Recognition Dr. Frank B. ter Haar TNO Defence, Security and Safety Oude Waalsdorperweg 63, P.O. Box 96864, 2509, JG The Hague We present an automatic and efficient method to fit a statistical deformation model of the human face to 3D scan data. In a global to local fitting scheme, the shape parameters of this model are optimized such that the produced instance of the model accurately fits the 3D scan data of the input face. To increase the expressiveness of the model and to produce a tighter fit of the model, our method fits a set of predefined face components and blends these components afterwards. Quantitative evaluation shows an improvement of the fitting results when multiple components are used instead of one. Compared to existing methods, our fully automatic method achieves a higher accuracy of the fitting results. The accurately generated face instances are manifold meshes without noise and holes, and can be effectively used for 3D face recognition. Our results show that model coefficient based face matching outperforms contour curve and landmark based face matching, and is more time efficient than contour curve matching.

Techniques described in this work apply to classes of shapes for which an average shape exists and other instances are variations of this average shape, such as the human face, the human body, and also 3D surface segmentations of the kidney, lung, and heart.


3


Object Recognition For Person and Car detection Rob Wijnhoven ViNotion BV / Technische Universiteit Eindhoven [email protected] / [email protected] Object detection is typically applied as detection-by-motion. Each image from the camera is compared with a model of the scene without objects. All pixels that are different from the background define moving objects. For constraint applications this is very well applicable, however, it cannot be applied when the camera is moving or the light conditions vary much. Furthermore, moving background objects like trees, flags and water cause false detections. In order to alleviate these problems, we present more advanced object models that model the appearance of objects and apply object detection by actually searching for similar objects in the image. Instead of analyzing motion, we apply detection-by-classification. These models are advantages, because they are insensitive to changes in light conditions, movement of the camera and background motion and can even be applied on single pictures (as compared to motion video). Furthermore, they can be applied in scenes with many moving objects and only detect objects of a single category. This is especially useful for specific scenarios like face detection, waterway surveillance and detection of pedestrians on highways. We have implemented the Histogram of Oriented Gradients (HOG) algorithm, a state-ofthe-art object detection algorithm and applied it onto some real-world practical datasets. Results are presented for the detection of persons for video-surveillance or consumer holiday pictures. Next, the same technique is applied for the detection of vehicles on panoramic pictures (cycloramas), taken while driving through a rural area. In the presentation, we will give an overview of the applied algorithm and present the challenges in the two mentioned applications. Furthermore, we introduce another application of object-detection for vessel analysis in the Rotterdam Harbor.


4


Examples of person-detection on consumer images

Examples of car-detection on images from a street


5


Text spotting in unconstrained environments Jochem van Vroonhoven Student University Utrecht, internship at Prime Vision [email protected] Textual information is all around us. It appears in many forms and at many places. Because text is very important for our communication, automatic text recognition in images and videos by a computer is a major research area since long. Previously this research was dealing with document images only. In recent years this was extended to natural scene images and video. Traditional text recognition approaches for document images do not produces reasonable results for the natural scene text problem. One of the reasons for this is that the text can appear anywhere in an image and at any orientation and scale. Our research has focused on a fast localization algorithm for text in natural scene video, which can be a part of a real-time scene text recognition system. Such a system is beneficial in many application areas, such as an aid to the visually impaired or robot navigation and guidance. The presented method consists of the following steps: preprocessing, segmentation, filtering, clustering and tracking. In the preprocessing step noise is reduced by a color clustering approach in HSV space. In the segmentation phase the image is binarized and segmented in connected components. These connected components are put through a filter based on edge and geometric features, to filter out most of the components that do not correspond to a text character. The remaining components are then clustered into lines of words, which are tracked over multiple frames. Evaluation of the proposed method on the image dataset of the ICDAR reading competitions has shown that the error rate of our method is comparable to other published methods, while the execution speed is much faster.


6


Real-time detection of surgical instruments in laparoscopic video: Towards automatic logging of surgical steps in the operating room Loubna Bouarfa1, Maja Rudinac1, Oytun Akman1, Jenny Dankelman, Pieter Jonker Department of Bio-mechanical Engineering, Delft University of Technology, Mekelweg 2, 2628CD,Delft, The Netherlands

Introduction The tracking of surgical instruments offers interesting possibilities for the development of an automatic logging system for the operating room. In a previous work [1,2], we showed that it is possible to recognize surgical steps from instrument signals with detection accuracies up to 90% using a labeled dataset. For this application it is important to automatically detect which instrument is used by the surgeon. If this problem is solved, the system can automatically log surgical steps online during surgery.

Pilot description Laparoscopic cholecystoctomy is one of the most frequent and standardized procedures in minimally invasive surgery. In this procedure four trocars are inserted to introduce laparoscopic instruments in the patient’s body.

Fig.1: Trocars inputs in Laparoscopic cholecystoctomy

Fig.1 illustrates the use of the trocars. The master trocar is used to insert active instruments, such as dissectors, scissors, and clip devices. The slave trocars are, however, used to insert graspers to hold the gallbladder for removal. As last, the view trocar is used to insert the endoscopic camera. Preliminary goal of this work is to detect whether or not an instrument is inserted into the appropriate trocars using laparoscopic video.

Algorithm description Due to the high noise level and varying illumination conditions image enhancement is performed as a preprocessing phase. As a first step Gaussian blurring is used for noise removal. Afterwards contrast enhancement and brightness normalization are performed for red color suppression. Finally specular reflections are removed to eliminate the overexposed regions. In the following processing phase, background is suppressed, instruments segmented using color segmentation in HS histogram and a binary image representing the possible instrument locations is generated.

1

Contributed equally to the work and therefore should be considered equivalent authors.


7


In the final phase, the binary image is post-processed using morphological operations. Hough lines are calculated and position and orientation of instruments are estimated.

Fig.2: Algorithm for automatic detection of laparoscopic instruments

Results & Conclusions:

Fig.3: Original Image

Fig.4: Instrument Segmentation

Fig.5: Detecting Instrument Orientation and Position

The algorithm was tested on videos from different segments of laparoscopic cholecystoctomy procedure. Early results show that the algorithm performs faster than real time and robust against surgery related distortions (e.g. smoke, blooding) (see Figures 3, 4, and 5).

References [1] L. Bouarfa, P.P. Jonker, J. Dankelman, high level task discovery in the operating room, to appear in the Journal of Biomedical Informatics (JBI), 2010 (accepted)


8


[2] L. Bouarfa, P.P. Jonker, J. Dankelman, Surgical context discovery by monitoring low-level activities in the OR, in Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2009


9


NHL Kenniscentrum Computer Vision Lab

Mei 2009

Computer Vision Computer Vision is het automatiseren van visuele inspecties. Met behulp van een computer worden beelden geïnterpreteerd die met een camera zijn vastgelegd. De op deze wijze verkregen informatie kan vervolgens worden gebruikt om andere processen aan te sturen. Voorbeelden daarvan zijn: - kwaliteitscontrole - positie- en oriëntatiebepaling - sorteren van producten op lopende banden. Voor veel bedrijven die producten maken of verwerken is visuele inspectie of meting belangrijk. Met behulp van Computer Vision is het in een groot aantal gevallen mogelijk om deze inspecties of metingen geautomatiseerd te laten uitvoeren. Dit zal in veel gevallen kunnen bijdragen aan een goedkoper, flexibeler en/of arbeidsvriendelijker productieproces. Computer Vision is een sectoroverschrijdende combinatie van Liveability en Creative Technologies, waar door toepassing van camera en display technologieën de kwaliteitscontrole van producten en diensten verbeterd kan worden.

Project “Inspectie scheerkappen” voor Philips Drachten

Mei 2009

Computer Vision Lab Het CV Lab van de NHL is in 1994 gestart en is inmiddels uitgegroeid tot een goed uitgerust laboratorium met een landelijke bekendheid. De missie van het CV Lab is: - Hét kenniscentrum worden op het gebied van Computer Vision voor het bedrijfsleven in Noord-Nederland - Hét kenniscentrum blijven op het gebied van Computer Vision voor het HBO in Nederland De doelstellingen van het CV Lab zijn: - Sectoroverschrijdende toepassing van technologie - Verrichten van contractactiviteiten en toegepast onderzoek - Verbreding en verdieping van kennis en expertise op het gebied van Computer Vision - Opbouwen kennisnetwerk en marktverkenning - Kennistransfer richting het bedrijfsleven - Verzorgen van onderwijs en PR instroom activiteiten - Integratie van de opleidingen W, E, TI en I door multidisciplinaire problemen De staf van het CV Lab bestaat uit een coördinator, een docent, een junior onderzoeker en twee projectingenieurs. Ruim 170 studenten hebben hun stage- of afstudeeropdracht bij het CV Lab uitgevoerd. De kracht van het CV Lab is kennis van en apparatuur voor de hele keten van: - belichting - camera’s - optiek - opstelling - algoritmen voor beeldverwerking - embedding in systemen of andere software

Projecten voor het bedrijfsleven Sinds 1996 zijn er meer dan 90 projecten voor het bedrijfsleven gestart en succesvol afgerond. Een typische projectomvang ligt tussen de 2.500,- en 40.000,- euro. Er is in totaal voor meer dan 1.800.000,- euro aan projecten uitgevoerd, vrijwel geheel toegevoegde waarde van NHL-uren toegepast onderzoek. Opdrachtgevers variëren van Friese eenmanszaken tot multinationals afkomstig uit geheel Nederland. Ongeveer de helft van de opdrachten zijn vervolgopdrachten.

Relatienetwerk Een uitgebreid relatienetwerk is opgebouwd door deelname als exposant op beurzen, het geven van lezingen en workshops, contacten met leveranciers, contacten met universiteiten en kennisinstituten, publicaties, het organiseren van ‘open’ dagen in het CV Lab en het geven van cursussen.

10



Mei 2009

Onderwijs Vanaf 1997 wordt het vak Computer Vision gegeven aan studenten van de NHL. Dit vak wordt nu gegeven bij het Instituut Techniek binnen de opleidingen Informatica, Technische Informatica, Elektrotechniek en Werktuigbouwkunde, én bij Life Science binnen de opleiding Forensic Science. De lesstof is in de loop der jaren geïnspireerd door ervaringen die zijn opgedaan bij de in opdracht van het bedrijfsleven uitgevoerde projecten. De lesstof wordt jaarlijks geactualiseerd en uitgebreid. De NHL heeft special voor dit vak een practicumzaal ingericht met daarin tien werkplekken met pc’s, camera’s en belichtingsapparatuur. De cursus wordt ook gegeven aan een Spaanse universiteit waarmee een uitwisselingsprogramma is opgestart. Ruim 30 studenten uit het buitenland hebben hun stage- of afstudeeropdracht bij het CV Lab uitgevoerd.

Practicumzaal NHL Kenniscentrum Computer Vision Lab

Cursus voor het bedrijfsleven Het vak Computer Vision is van 11 tot en met 15 mei 2009 voor de twaalfde keer gegeven als cursus van een week voor mensen uit het bedrijfsleven. Platform Beeldverwerking HBO Op verzoek van Senter Novem IOP Beeldverwerking (Economische Zaken) heeft het CV Lab een plan opgesteld om de kennis op het gebied van beeldverwerking te verankeren bij het HBO en MKB in Nederland. Dit heeft geleid tot toetreding als lid van de begeleidingscommissie van de IOP Beeldverwerking en de oprichting van het Platform Beeldverwerking voor het HBO. In dit platform vervult de NHL een voortrekkersrol. Negen andere hogescholen hebben zich inmiddels aangesloten bij dit platform. Een van de activiteiten van dit platform is het opleiden van HBO docenten tot docent Computer Vision. Meer dan vijfentwintig HBO docenten volgden daartoe de vijfdaagse cursus Computer Vision van de NHL. De rechten op deze cursus zijn aan de andere hogescholen verkocht voor gebruik ten behoeve van onderwijs aan hun reguliere studenten.

Mei 2009

Cluster Computer Vision Noord Nederland Het CV Lab heeft de Cluster Computer Vision Noord Nederland opgericht. Dit is een platform voor bedrijven in Noord-Nederland die zich bezig houden met computer vision. Inmiddels hebben zich hiervoor 29 bedrijven gemeld uit Noord-Holland, Friesland, Groningen, Drenthe en Overijssel. De doelstellingen van de Cluster Computer Vision Noord-Nederland zijn: - het aanspreekpunt voor Computer Vision in Noord Nederland worden - het bevorderen en structuren van de kenniscirculatie op het gebied van computer vision in Noord-Nederland onder de volgende partijen: - het bedrijfsleven (vooral ook MKB) - de kennisinstituten (universiteiten, HBO’s, TNO etc) - de intermédiaire organisaties (Syntens, MKB, VNO etc.) - het hebben van een netwerkfunctie - het stimuleren van samenwerkingsprojecten bij de deelnemers, bijvoorbeeld onderzoek of productontwikkeling het promoten van de toepassing van computer vision bij bedrijven -

Robo Challenge De Robo Challenge is een jaarlijks terugkerende wedstrijd voor autonoom voorbewegende robots, die gekleurde ballen die aan draden hangen moet verzamelen. Aan deze wedstrijd wordt deelgenomen door teams van hogescholen, universiteiten, bedrijven en privé personen. In 2007 en 2008 is deze wedstrijd door de NHL gewonnen.

11



Mei 2009

Toekomst De vraag van het bedrijfsleven naar automatisering van visuele inspecties neemt de laatste jaren sterk toe. Omdat kwaliteitscontrole van producten steeds belangrijker wordt is de verwachting dat deze groei zal doorzetten. Kwaliteitscontrole m.b.v. Computer Vision is vaak sneller en betrouwbaarder dan het menselijk oog. Bij veel bedrijven zijn de mogelijkheden van Computer Vision nog onbekend, en bedrijven die ermee gestart zijn zien steeds meer toepassingsmogelijkheden. Het CV Lab ziet hier een belangrijke rol om de kennis m.b.t. Computer Vision uit te dragen bij zowel het bedrijfsleven als het HBO. Het CV Lab kan door het ontwikkelen van toegepast onderzoek een brugfunctie vervullen tussen het fundamenteel onderzoek van universiteiten en de toepassingen van Computer Vision in het bedrijfsleven. Het CV Lab heeft samenwerkingsverbanden opgestart die verder uitgebouwd kunnen worden in nieuwe sectoroverschrijdende toepassingen zoals: - Forensic science (o.a. Rapid DNA Detection) - Life science - Integrale Veiligheid - Serious Gaming - Multimedia - E-care en Domotica Zo kan de toepassing van Computer Vision veel breder worden dan de huidige toepassing binnen met name de sector van de Industriële Automatisering. De komende jaren ziet het CV Lab de volgende nieuwe technische speerpunten: - Classificatie m.b.v. zelflerende systemen (neurale netwerken) - Thermal Imaging (infrarood camera’s) - Vision en Robotica - Range imaging

Voor meer informatie: Jaap van de Loosdrecht Coordinator NHL Kenniscentrum Computer Vision Tesselschadestraat 12, 8913 HB Leeuwarden telefoonnummer: 058 251 1193, mobiel: 06 139 49 207, fax: 058 251 1115 [email protected] www.nhl.nl/computervision

12



Elementary Multi-scale Riemann-Finsler geometry Applications to Tensor Valued Medical Image Analysis Laura Astola Application

Regularization of Noise

High Angular Resolution Diffusion Imaging (HARDI) is a collective name for medical imaging techniques that uses diffusion weighted NMR-measurements to infer the thermal motion of water molecules within human tissue.

Noisy diffusion profiles are regularized by solving the heat equation on the sphere. (

D(0) D(τ )

P = N Di1 ···in yi1 · · · yin Pn=0 N = n=0 eτ n(n+1) Di1 ···in yi1 · · · yin .

This can be done faster without iterative fittings: • Solve initial tensor d = MT M

Figure 1. Instead of a scalar value, multiple scalars are measured in different directions. A smooth surface is fitted to each set of measurement.

−1

MT m.

• Compute the regularized tensor Ad, where A is a precomputed matrix containing regularization coefficients depending only on the degree of the tensor, constructed using Clebsch projection, Kronecker product and symmetrization.

Finsler Model Fit an even order tensor Dn to data minimizing L2 distance and define a diffusion profile F (y) as F (y) = Di1 ···in yi1 · · · y where

1 in n

With a Finsler metric we can implement streamline tracking to track candidates for axon bundles.

,  

y = (sin θ cos ϕ, sin θ sin ϕ, cos θ) .



Then for each pair (x, y) we obtain a ”diffusion tensor” Dij (y) =

Fiber Tracking with Finsler Metric

c˙ = arg max Dij v j |v|=1

c(0) = p .

1 ∂2F 2 . 2 ∂yi ∂yj

Figure 3. Left: Fibers tracked in a simulated crossing. Right: Fibers tracked in real data showing crossings of the corona radiata and the corpus callosum.

Figure 2. Left: Tensor fitted to data. Right: Each vector determines a metric tensor.


13

/centre for analysis, scientific computing and applications


Dual tensor atlas generation based on a cohort of coregistered non-HARDI datasets Matthan W.A. Caan, C.A. Sage, M.M. van der Graaf, C.A. Grimbergen, S. Sunaert, L.J. van Vliet and F.M. Vos Department of Radiology, Academic Medical Center, University of Amsterdam, NL Quantitative Imaging Group, Delft University of Technology, NL Dept of Radiology, University Hospitals of the Catholic University of Leuven, BE This work has been published and presented on MICCAI'09, September, in London.

We propose a method to create a dual tensor atlas from multiple coregistered MR Diffusion Weighted datasets, without High Angular Resolution Diffusion Imaging (nonHARDI). Increased angular resolution is ensured by random variations of subject positioning in the scanner and different local rotations applied during coregistration resulting in dispersed gradient directions. Simulations incorporating residual coregistration misalignments show that using 10 subjects should already double the angular resolution, even at a relatively low b-value of b=1000 s/mm2. Commissural corpus callosum fibers reconstructed by our method closely approximated those found in a High Angular Resolution Diffusion Imaging (HARDI)-dataset.

Tracked corpus callosal fibers (green) and commissural fibers (magenta) in our dual tensor atlas. The latter cannot be tracked using a single tensor model


14


Tractography in anatomical images of wrist and knee: regularization of ligament orientation estimates Martijn van de Giessen1,2, Frans M. Vos1,3, Laurens den Haan1, Kees A. Grimbergen2, Lucas J. van Vliet1, Geert J. Streekstra2 Delft University of Technology, Faculty of Applied Sciences, Quantitative Imaging Group1 Lorentzweg 1, 2628 CJ Delft, The Netherlands Academic Medical Center, Depts. Biomedical Engineering and Physics2, Radiology3 Meibergdreef 9, 1105 AZ Amsterdam, The Netherlands

The analysis of ligament trajectories is currently limited to qualitative anatomical maps, e.g. [1], measurements of lengths and thicknesses or orientation studies in two dimensional slices. For better insight in the functioning of joints and the development of (cruciate) ligament replacements, a quantitative knowledge of the three-dimensional collagen fiber trajectories in ligaments is needed, as the forces in ligaments are directed along these trajectories. Orientation measurements are at the basis of estimating ligament trajectories, in this work in anatomical images. Current orientation measurements are unreliable in the neighbourhood of crossing bundles and near ligament-bone insertions. Two non-linear methods are proposed for regularization of orientation estimates in these neighbourhoods, based on the generalized Kuwahara filter [2] and clustering of orientations [3]. The two non-linear regularization methods give improved results over an existing method based on normalized convolution [4] in both simulations and anatomical images of ligaments in wrist and knee data. Finally, a tractography similar as in diffusion tensor imaging data is used to measure trajectories of ligaments in the wrist and knee (See the figure for tracts in the wrist).

Normalized convolution

Generalized Kuwahara

Orientation clustering

Projection of tracts in two crossing ligamentous structures (a and b) in the wrist. Tracts go in two directions and start at the black points. [1] Richard A. Berger, “The ligaments of the wrist: A current overview of anatomy with considerations of their potential functions," Hand Clin., vol. 13, pp. 63-82, 1997. [2] P. Bakker, L. J. Van Vliet, and P. W. Verbeek, “Edge preserving orientation adaptive filtering," in CVPR, 1999, 1999, vol. 1, p. 540 Vol. 1. [3] L. J. Van Vliet and F. G. A. Faas, “Multi-orientation analysis by decomposing the structure tensor and clustering," in ICPR, 2006, 2006, vol. 3, pp. 856-860. [4] C. F. Westin and H. Knutsson, “Tensor field regularization using normalized convolution," EUROCAST 2003, LNCS 2809, pp. 564–572, 2003.


15


DNA Deformations near Charged Surfaces: Electron and Atomic Force Microscopy Views F. G. A. Faas1, B. Rieger1, L. J. van Vliet1, D. I. Cherny2 1 Department of Imaging Science and Technology Delft University of Technology, Delft, The Netherlands; 2 Department of Biochemistry, University of Leicester, Leicester, United Kingdom [email protected]

DNA is a very important cell structural element, which determines the level of expression of genes by virtue of its interaction with regulatory proteins. We use electron (EM) and atomic force microscopy (AFM) to characterize the flexibility of double-stranded DNA (~150–950 nm long) close to a charged surface. Automated procedures for the extraction of DNA contours combined with new statistical chain descriptors indicate a uniquely 2D equilibration of the molecules on the substrate surface regardless of the procedure of molecule mounting. However, in contrast to AFM, the EM mounting leads to a noticeable decrease in DNA persistence length together with decreased kurtosis. Analysis of local bending on short length scales (down to 6 nm in the EM study) shows that DNA flexibility behaves as predicted by the wormlike chain model in contrast to recent other findings. We therefore argue that adhesion of DNA to a charged surface may lead to additional static bending (kinking) of ~5 degrees per dinucleotide step without impairing the dynamic behavior of the DNA backbone. We use an earlier developed improved Fast Marching algorithm after automatic segmentation of the molecules to find the DNA centerline representing the backbone. The image analysis algorithm was extensively validated on images generated by Monte Carlo simulations. In the data analysis, extra precaution was taken in sampling the DNA strands to avoid correlation. Reusing all data for computing the characteristic quantities for each length along the DNA yield highly correlated points. To avoid this, we divided each DNA strand into length segments randomly drawn from a predefined set of lengths such that no piece of strand is used twice.


16


Correspondence free 3D statistical shape model fitting to sparse X-ray projections N. Bakaa,b, W.J. Niessenb, B.L.Kapteinc, T. van Walsumb, L. Ferrarinia, J.H.C. Reibera, B.P.F. Lelieveldta a) Division of Image Processing, Department of Radiology, Leiden University Medical Center (LUMC), Leiden b) Biomedical Imaging Group Rotterdam, Erasmus Medical Center, Rotterdam c) Department of Orthopaedic Surgery, LUMC, Leiden [email protected] In the presentation we address the problem of 3D shape reconstruction from sparse X-ray projections. We present a correspondence free method to fit a statistical shape model to two X-ray projections, and illustrate its performance in 3D shape reconstruction of the femur. The method alternates between 2D segmentation and 3D reconstruction, where 2D segmentation is guided by dynamic programming along the model projection on the Xray plane. 3D reconstruction is based on the iterative minimization of the 3D distance between a set of support points and the back-projected silhouette with respect to the pose and model parameters. We show robustness of the reconstruction on simulated X-ray projection data of the femur, varying the field of view; and in a pilot study on cadaveric femora.

Illustration of the support point selection for 3D shape reconstruction


17


Integrated visualization and analysis of coronary arteries and myocardial perfusion H.A.Kirisli, Department of Medical Informatics & Radiology, Erasmus MC, Rotterdam Nowadays, multiple imaging modalities are available to diagnose cardiovascular heart disease. We present a visualization tool that significantly improves the diagnosis of heart disease by relating CTA data to perfusion MRI data, using a patient-specific approach. Cardiac magnetic resonance imaging (MRI) is widely used to assess myocardial perfusion and viability, allowing detection of ischemic myocardial disease. Computed tomography angiography (CTA), a non-invasive technique, is becoming increasingly popular for the diagnosis of coronary artery disease (CAD), instead of the conventional and invasive coronary angiography modality. This increased usage of CTA for cardiac examinations is mainly due to its superior 3D spatial and temporal resolution. Myocardial perfusion and viability information derived from MRI are visualized with a standardized so-called Bull’s Eye Plot (BEP), while coronary arteries are often visualized using curved planar reformatted images of the 3D CTA. Assumptions on coronary artery territories combined with a mental integration of myocardial function information and coronary arteries anatomy leads up to a diagnosis. However, it is a fastidious task for clinicians, and the anatomical variability of the coronary arteries among individuals is very high. Our tool enables the fusion of 3D models of the coronary artery anatomy with myocardial physiology information, and facilitates integrated visualization and analysis. A spatial correspondence is directly established between diseased coronary artery segments and myocardial regions with abnormal perfusion. This allows the location of coronary stenoses and perfusion abnormalities to be visualized jointly in 3D, thereby facilitating the study of the relation between the anatomic causes of a blocked artery and the physiological effects on the myocardium perfusion. It is a useful tool for the assessment of atherosclerosis, and it may open new areas for diagnosis, prevention, and treatment of coronary atherosclerosis.

Fig.1 – CTA coronary arteries overlaid on perfusion BEP

Fig.2 CTA coronary arteries overlaid on perfusion BEP, with 3D coronary tree


18


Fig.3 –Perfusion information overlaid on CTA data


19


Coronary intervention planning by fusing angiogram and IVUS Shengxian Tu1, Gerhard Koning1, Zheng Huang2, Kai Cui2, Andrei Rares1, Jasper P. Janssen1, Johan H. C. Reiber1 1 LKEB, Dept. Radiology, Leiden University Medical Center, Netherlands 2 Dept. Cardiology, Nanfang Hospital, China [email protected] Coronary angioplasty is a minimally invasive procedure to open obstructed arteries under the guidance of X-ray angiography. Despite of the tremendous success that has been made by the procedure in the instant treatment of coronary artery diseases, high restenosis rate due to the suboptimal stent deployment has hampered the translation of the procedure success into longterm outcomes. Dedicated intervention planning could be extremely helpful in optimizing the procedure to reduce the restenosis rate. A novel system by fusing angiographic data and intravascular ultrasound (IVUS) has been developed to increase the reliability of the interpreted dimensions about the diseased vessel segments. Based on two 2D angiographic views and the geometric data of the acquisitions, an accurate and robust 3D reconstruction of the vessel segments of interest can be achieved by first correcting for the system distortion introduced by the isocenter offset and next by the reconstruction of vessel centerline and cross sections. In the next step the reconstructed segments are registered with IVUS catheter path by a distance mapping algorithm to correspond angiographic and IVUS data. The registration allows to interpreting plaque characteristics with accurate geometric properties for the diseased segments as well as recommending optimal viewing angles for the online intervention, by which the stent deployment is expected to be optimized, especially for bifurcation intervention. The 3D angiographic reconstruction was validated to be of high accuracy and precision by phantoms experiments: Accuracy and precision for segment length assessment were 0.04 ± 0.25 mm (P < 0.01), for optimal viewing angle assessment were -1.5º ± 3.6º (P < 0.01) and -0.2º ± 2.4º (P = 0.54), in term of rotation angle and angulation angle, respectively. In-vivo fusion of angiographic and IVUS data showed the feasibility of combining morphological and functional information for appropriate stent selection and deployment.


20

NVPHBV - Fall Meeting 2009

Recommend Documents