Fuzzy-Based Extraction of Vascular Structures from Time-of-Flight MR - PowerPoint PPT Presentation

Department of Medical Informatics Fuzzy-Based Extraction of Vascular Structures from Time-of-Flight MR Images N.D. Forkert 1 , D. Säring 1 , K. Wenzel 2 , T. Illies 2 , J. Fiehler 2 , H. Handels 1 1 Department of Medical Informatics 2 Department of Diagnostic and Interventional Neuroradiology University Medical Center Hamburg-Eppendorf

Department of Medical Informatics Cerebral Vascular Diseases • ! The stroke is third most common reason for death in Europe • ! 80% are caused by a ischemia (lack of blood supply) • ! In 20 % are caused by hemorrhage • ! Hemorrhages are mostly caused by a rupture of a malformed blood vessel. • ! Examples for malformations: – ! Aneurysms – ! Arteriovenous malformations (AVM)

Department of Medical Informatics Therapy of Cerebral Vessel Malformations • ! Neurosurgical Resection • ! Endovascular Treatment – ! Embolisation – ! Coiling • ! Radiosurgery • ! Decision of treatment mode(s) – ! Size – ! Location – ! … Exact knowledge about the individual Image of AVM surgery anatomy of the vascular system is needed for risk estimation and therapy

Department of Medical Informatics 3D-Time-of-Flight (TOF) MRA • ! 3D Time-of-Flight technique is often used for visualization of the vascular system – ! 3T MR Scanner (Siemens) - ! Matrix: 384 x 512 * - ! Size: 0.47 x 0.47 mm * (* = typical values) - ! Slice thickness: 0.5 mm - ! 156 Slices - ! Anatomical Information • ! Improved blood to background contrast 3D TOF-MRA with improved blood to background contrast

Department of Medical Informatics Artifact Reduction • ! Multi-Slab technology acquisition of TOF image sequences  ! Reduction of the amplitude in overlapping regions MIP-Visualization of a TOF image sequence (Slab Boundary Artifact) • ! Reduction of the Slab Boundary Artifact using histogram matching* MIP-Visualization of a TOF image sequence after Slab Boundary Artifact reduction * Kholmovski et al. Correction of Slab Boundary Artifact Using Histogramm Matching. J Magn Reson Imaging. 2002;15:610-617.

Department of Medical Informatics Skull Stripping • ! Drawback of TOF image data: Non-cerebral tissues like fat, bone marrow and eyes are represented by intensities similar to the vessels  ! hindered segmentation of blood vessels • ! Skull Stripping using a graph-based approach* Volume Rendered TOF image before (top) and after (bottom) skull stripping *Forkert et al. Automatic Brain Segmentation in Time-of-Flight MRA Images Methods of Information in Medicine, 48(5), 2009 (in press)

Department of Medical Informatics Vessel Segmentation – State of the Art • ! Intensity-based approaches: – ! Typical intensity distributions – ! e.g. Z-Buffer-Segmentation 1 (global threshold) – ! Problem: small vessels are often not detected • ! Model-based approaches: – ! Typical vessel morpholgy – ! e.g. Vesselness-Filter 2 – ! Problem: vessel malformations are often not detected Vesselness-Result TOF-Slice → Combination of intensity- and shape-information 1 Chapman et al., Intracranial vessel segmentation from time-of-flight MRA using preprocessing of the MIP Z-Buffer, Medical Image Analysis 8 (2004), 113-126. 2 Sato et al. Three-dimensional multi-scale Line Filter for Segmentation and Visualization of Curvelinear Structures in Medical Images. Med Image Anal. 1998;2(2):143-168.

Department of Medical Informatics Fuzzy Vessel Segmentation Preprocessing 1. Preprocessing 2. Fuzzy Inference System 3. Fuzzy Extraction Fuzzy Inference System Fuzzy- Connectedness

Department of Medical Informatics 1. Preprocessing • ! Computation of the Vesselness- Image* – ! A value of vesselness measure is assigned to every voxel based on eigen values of the Hessian matrix – ! Benefit: Enhanced display of the TOF-Slice vascular structures, especially small vessels – ! Drawback: Vascular malformations are not detected • ! Computation of the Maximum- Image – ! The maximal intensity within a defined 3D neighborhood is assigned to every Vesselness- Maximum- Image Image voxel *Sato et al. Three-dimensional multi-scale Line Filter for Segmentation and Visualization of Curvelinear Structures in Medical Images. Med Image Anal. 1998;2(2):143-168.

Department of Medical Informatics 2. Fuzzy Inference System • ! Combination of the images using fuzzy Inference system • ! Benefits: – ! Non-linear combination parameters – ! Inclusion of uncertain knowledge – ! Well explored and broad utilization in control engineering • ! Steps: – ! Fuzzyfication – ! Inferenz » ! Aggregation » ! Implication » ! Accumulation – ! Defuzzyfication Fuzzy Inference System

Department of Medical Informatics 2. Fuzzy Inference System • ! Fuzzyfication: – ! Sharp input value specifies a degree of membership of each fuzzy set – ! 3 Fuzzy-Sets (low, medium, high) – ! Functions for Fuzzy-sets are generated automatically based on empirical knowledge – ! Example: Intensity value 210 leads to a degree of 0.8 to medium and 0.2 to high

Department of Medical Informatics 2. Fuzzy Inference System • ! Fuzzy Inference: – ! 27 rules (3 inputs with 3 linguistic terms) – ! 5 conclusions for vessel probability: very low, low, medium, high, very high – ! Main idea for definition of the rule base: Weight the response of the vesselness filter stronger if the maximum filter responds a low value, whereas the TOF-input is weighted stronger otherwise .

Department of Medical Informatics 2. Fuzzy Inference System • ! Example for a rule: If the TOF-Intensity is „medium“ and the vesselness-measure „high“ and the Maximum-Filter-value „low“ then the vessel probability is „high“. • ! Aggregation: – ! combination the degrees of membership of the premise parts of a rule to one value for the whole premise – ! Minimum-Operators – ! Example: » ! 0.9 for “TOF Intensity is medium” » ! 0.7 for “Vesselness-Measure is high” » ! 0.7 for “Maximumfilter-Value is low” – ! Degree for the premise: min(0.9, 0.7, 0.7) = 0.7

Department of Medical Informatics 2. Fuzzy Inference System • ! Example for a rule: If the TOF-Intensity is „medium“ and the vesselness-measure „high“ and the Maximum-Filter-value „low“ then the vessel probability is „high“. • ! Aggregation • ! Implication: – ! Determinition of the membership degree of the conclusion – ! cutting the fuzzy set

Department of Medical Informatics 2. Fuzzy Inference System • ! Example for a rule: If the TOF-Intensity is „medium“ and the vesselness-measure „high“ and the Maximum-Filter-value „low“ then the vessel probability is „high“. • ! Aggregation • ! Implication • ! Accumulation: – ! Accumulation of the single results of all rules – ! Maximum-Operator

Department of Medical Informatics 2. Fuzzy Inference System • ! Example for a rule: If the TOF-Intensity is „medium“ and the vesselness-measure „high“ and the Maximum-Filter-value „low“ then the vessel probability is „high“. • ! Aggregation • ! Implication • ! Accumulation • ! Defuzzyfication – ! Calculation of a sharp output value – ! Center of gravity method

Department of Medical Informatics 3. Fuzzy Extraction • ! Fuzzy-Parameter-Image: – ! Small Vessels as well as malformation are enhanced • ! Vessel extraction: – ! Global thresholding – ! Connected component analysis – ! Mean and standard deviation computation of the fuzzy values of each component – ! Fuzzy-Connectedness Approach* – ! Result: Extracted vascular system Fuzzy-Parameter-Image *Udupa et al.: Fuzzy Connectedness and Object Denition: Theory, Algorithms, and Applications in Image Segmentation. Graphical Models. 1996;58(3):246-261.

Department of Medical Informatics Results • ! 17 datasets of patients with an arteriovenous malformation • ! Manual segmentation: – ! Semi-automatic segmentation – ! Volume-Growing – ! Manual correction in orthogonal views – ! Time requirements: 8-12 hours – ! Performed by neuroradiologists • ! Automatic segmentations: – ! Z-Buffer Segmentation, Time: ~5 min – ! Fuzzy Segmentation: ~30 min • ! Evaluation – ! Dice-Value – ! Kappa-Value

Department of Medical Informatics Results • ! Mean Dice-Value – ! Z-Buffer Segmentation: 0.577 – ! Fuzzy Segmentation: 0.742 • ! Mean Kappa-Value – ! Z-Buffer Segmentation: 0,567 – ! Fuzzy Segmentation: 0,775

Department of Medical Informatics Results • ! The quantitative results depend on the size of the AVM-nidus • ! One dataset with two manual segmentations from different experts  ! Inter-Observer Comparison Dice-values in dependency to the nidus size

Department of Medical Informatics Results (Inter-Observer-Comparison) • ! Mean Dice-Value – ! Inter-Observer Comparison: 0,83 – ! Z-Buffer Segmentation: 0.695 – ! Fuzzy Segmentation: 0,815 • ! Mean Kappa-Value – ! Inter-Observer Comparison: 0.84 – ! Z-Buffer Segmentation: 0.755 – ! Fuzzy Segmentation: 0,845

Department of Medical Informatics Results • ! Section of a TOF slice Manual Segmentations Z-Buffer Fuzzy Segmentation Segmentation

Department of Medical Informatics Results • ! Surface Models Manual segmentation Z-Buffer segmentation Fuzzy segmentation