230 likes | 338 Views
( Adaptive ) enhancement of HARDI images. Eric Creusen Supervisor: Remco Duits. Diffusion of Water. Diffusion is dependent on orientation. * Graphic borrowed from Thomas Schultz. Diffusion of water: Modelling. Domain.
E N D
(Adaptive) enhancement of HARDI images Eric Creusen Supervisor: Remco Duits
Diffusion of Water Diffusion is dependent on orientation *Graphicborrowedfrom Thomas Schultz / BIOMIM
Domain Diffusionweighted images are functionsonposition and orientation Both HARDI and DTI! / Communicatie Expertise Centrum
Embed in group Embedposition and orientationintotranslations and rotationgroup / BIOMIM
Parametrization of Sphere Euler Angles / Communicatie Expertise Centrum
Embeddingproblem SE(3) has 3 angles(α,β,γ) Positions and orientationonly 2 (β,γ) Orientations are independent of α Introduceequivalence classes on SE(3) Processesshould preserve α-invariance! / BIOMIM
Convection-diffusionprocesses Seefunction as distribution of orientedparticles Particlescan translate and rotate / BIOMIM
Rotatedcoordinate system A1 z A3 n n Rn (α) y A2 x / Communicatie Expertise Centrum
Now that we have a basis: derivatives! In the same direction as normal Angular derivatives Should be zero Perpendicular (make random choice here) Derivativescanbecalculated Directionsdependantonorientationn 3 spatialdirections 2 angulardirections (+1 dummy direction) For n=ez / Communicatie Expertise Centrum
Smoothing processes Withderivatives, we cansimulatediffusionprocessesbysimulatingPDE’s General diffusionprocesses are givenby the PDE: (with parameter D and simulation time t) / Communicatie Expertise Centrum
Some simple examples Notinteresting, onlyspatial OR onlyangulardiffusion • IsotropicSpatialGaussian smoothing: • Angulardiffusion: / Communicatie Expertise Centrum
More examples “Contour enhancement” Spatial diffusion Angular diffusion “Contour completion” Convection term / Communicatie Expertise Centrum
Simulation of PDEs • Convolutions with kernels • Finite difference methods / Communicatie Expertise Centrum
Creating convolution kernels Getapproximationswithcomplicatedmath……. OR Start withoriented delta peak Simulate PDE withfinitedifferencemethods / Communicatie Expertise Centrum
Finite difference methods: enhancement enhancement Directly use finite difference methods to simulate PDE / Communicatie Expertise Centrum
Finite difference: contour completion completion / Communicatie Expertise Centrum
Adaptive filtering Makesmoothingdependanton data Finitedifferencemethodsneccesary / Communicatie Expertise Centrum
Perona-Malik type diffusion “Contour enhancement” Goal: Edge preserving smoothing Contour enhancement, but stop diffusionacrossedges / BIOMIM
Future work • Adaptivediffusionprocesses • Usecurvature and torsionforadaptiveprocesses • Combine diffusion and erosionprocesses • Start lookinginto practical applications / Communicatie Expertise Centrum
Questions? / Communicatie Expertise Centrum