medical image analysis -
TRANSCRIPT
1
Medical Image Analysis
Grégoire Malandain
Application to 3D angiography
16/01/2003 Grégoire Malandain 2
2D Visualisation
Adapted to large volumes
2
16/01/2003 Grégoire Malandain 3
2D Visualisation
Adapted to large volumes
… but not to vessels
16/01/2003 Grégoire Malandain 4
Maximum Intensity Projection
+ Uses original intensity values
– Does not allow 3D perception
+ but allowed by animation
– does not allow quantification
3
16/01/2003 Grégoire Malandain 5
Isosurfaces+ It’s a segmentation
+ Allows quantification
– Effectiveness?
Threshold = 14
16/01/2003 Grégoire Malandain 6
Threshold = 29
Isosurfaces+ It’s a segmentation
+ Allows quantification
– Effectiveness?
4
16/01/2003 Grégoire Malandain 7
Threshold = 44
Isosurfaces+ It’s a segmentation
+ Allows quantification
– Effectiveness?
16/01/2003 Grégoire Malandain 8
Isosurfaces– Effectiveness of segmentation?
Topology changes
5
16/01/2003 Grégoire Malandain 9
Segmentation issues
• Geometrical correctnessMeasure of length, diameter, area, ...
• Topological correctnessExtraction of the central line
Retrieval of junctionsSeparation of adjacent structures
16/01/2003 Grégoire Malandain 10
Principles
• Vessel modelling
• Vessel characterisation
• Central line extraction
• Radius estimation
6
16/01/2003 Grégoire Malandain 11
3D vessels modelling
• Tubular-like structures• Shape of the section?
• Intensity distribution inside the section?
Gaussian Constant Gaussian * Constant
16/01/2003 Grégoire Malandain 12
Intensity profile: 1D derivatives)2/-(OM 22
e)( RCMI =
7
16/01/2003 Grégoire Malandain 13
Intensity profile: 1D derivatives)2/-(OM 22
e)( RCMI =
xI
∂∂
16/01/2003 Grégoire Malandain 14
Intensity profile: 1D derivatives)2/-(OM 22
e)( RCMI =
2
2
xI
∂∂
8
16/01/2003 Grégoire Malandain 15
Intensity profile: 1D derivatives)2/-(OM 22
e)( RCMI =
)()()( rMxIrM
xIMResp +
∂∂−−
∂∂=
Maximal response
)()( ε±>⇔ MRespMResp
16/01/2003 Grégoire Malandain 16
• Shape characterisation (white structure on dark background):
Intensity profile: 3D derivatives
• Matrix of second derivatives: Hessian
IH
=
zzyzxz
yzyyxy
xzxyxx
IIIIIIIII
[ ]
=
3
2
1
3
2
1
321
VVV
VVVT
T
T
λλ
λ
iivvIi =λ
Sheet-like structure: 0, 0 |||,||| 321321 ≈<>> λλλλλλ
0 0, |||||,| 321321 ≈<>> λλλλλλTubular structure:
Cylinder direction: 3Vr
9
16/01/2003 Grégoire Malandain 17
Intensity profile: 3D response
• Vector of first derivatives: gradient
=∇
z
y
x
III
I
( )∫ +∇−=π
α2
0
v).v()( drMIMRrrr
Cylinder direction: 3Vr
Directions in the orthogonal section: 21 V)sin(V)cos(Vrrr
αα +=
• 3D response
Maximal response 1,2i ),v()( i =±>⇔r
εMRMR
16/01/2003 Grégoire Malandain 18
Derivatives computation• Discrete signal: 3D lattice of voxels
M(i,j,k)=I(x,y,z)
),,( zyxIx∂
∂?
• Convolution with a low-pass filter( ) I
xGzyxIzyxG
xzyxI
x*),,(*),,(),,(
∂∂=
∂∂≈
∂∂
10
16/01/2003 Grégoire Malandain 19
Practical issues
• Vessel of unknown radius σ0Gaussian intensity profiles
• Derivatives computed with a Gaussian σResponse computed at a distance θσ
How to compare scales?
• Multi-scale extraction: several σ’s
16/01/2003 Grégoire Malandain 20
Multiscale issues
• Comparing scalesγ-derivatives:
• Response extrema in space and in scale
• Optimal value of θ
• Estimation of σ0
xx ∂=∂ γγ σ,
• The profile modelling by a Gaussian
invariance w.r.t. scaling: γ=1
11
16/01/2003 Grégoire Malandain 21
σ=1.0 σ=1.28 σ=1.65
σ=2.12 σ=2.72 σ=3.5
16/01/2003 Grégoire Malandain 22
Segmentation issues
Original image Multi-scale response Multi-scale extrema
• How to recognise curves in the extrema?
12
16/01/2003 Grégoire Malandain 23
3-D Digital Topology
16/01/2003 Grégoire Malandain 24
Number of (26-)connected components of the foreground
in the (26-)neighborhood
3-D Digital Topology
C
*C
Number of (6-)connected components of the background
in the (18-)neighborhood, (6-adjacent to the centre)
13
16/01/2003 Grégoire Malandain 25
3-D Digital Topology
• Thinning the extrema( )1* == CCIterative removal of simple points
• Removing small pieces of curve
( )1 , 2* == CC( )1 , 2* => CC
Simple curves
are ended either by curves junctions
or borders (i.e. simple points)
16/01/2003 Grégoire Malandain 26
Original image Multiscale response Multiscale extremaPruned extrema
14
16/01/2003 Grégoire Malandain 27
Reconstructions
• Simple curves smoothing
• ReconstructionRadius estimation from the maximal scale
Intensity profile from the vessel modelling
Advantages of both MIP and isosurfaces
16/01/2003 Grégoire Malandain 28
About radius estimation
LW = 2 mm E = 4 mm
SD = 2 mm LD = 4 mmVoxel = 0.267 mm3
15
16/01/2003 Grégoire Malandain 29
About radius estimation
• EstimationLW: 2.018 mm E: 3.884-3.968 mm
SD: 2.14 mm LD: 3.6 mm
• is disturbed by 3-D computation
16/01/2003 Grégoire Malandain 30
About radius estimation
• Estimation in 2D resampled sections
• Better but computationally more expensive
16
16/01/2003 Grégoire Malandain 31
2D orthogonal sections resampling
• Subvolume
16/01/2003 Grégoire Malandain 32
2D orthogonal sections resampling
• Subvolume (scanner)
17
16/01/2003 Grégoire Malandain 33
2D orthogonal sections resampling
• Subvolume (MRA)
16/01/2003 Grégoire Malandain 34
2-D cross-sections: MR image
18
16/01/2003 Grégoire Malandain 35
MR image: extrema computation
axial coronal sagittal
16/01/2003 Grégoire Malandain 36
MR image: 2-D cross-sectionsResampling withcubic splines
Travelling inside thevessel
19
16/01/2003 Grégoire Malandain 37
2-D cross-sections: segmentation3 Thresholding: iso-contours
3 Edges detection
å zero-crossings
3 Sub-pixel accuracy
0=∆I0.. =∇∇ IHII
16/01/2003 Grégoire Malandain 38
2-D cross-sections: segmentation
20
16/01/2003 Grégoire Malandain 39
2-D cross-sections: segmentation
3 allows to build a 3-D mesh
(computational geometry, Prismeteam)
3 numerical simulation
16/01/2003 Grégoire Malandain 40
Synthetical model
Retour menu
Known vessels radii
21
16/01/2003 Grégoire Malandain 41
2-D cross-sections: segmentation
Diameter = 2.0 mm
derivatives with gaussian ofσ = 0.5 voxel
0.. =∇∇ IHII
16/01/2003 Grégoire Malandain 42
2-D cross-sections: segmentation
Diameter < 1.0 mm
derivatives with gaussian ofσ = 0.5 voxel
0.. =∇∇ IHII
22
16/01/2003 Grégoire Malandain 43
2-D cross-sections: segmentation
Diameter = 2.7 mm
derivatives with gaussian ofσ = 0.5 voxel
0.. =∇∇ IHII
16/01/2003 Grégoire Malandain 44
2-D cross-sections: segmentation
Diameter = 2.7 mm
derivatives with gaussian ofσ = 1.0 voxel
0.. =∇∇ IHII
23
16/01/2003 Grégoire Malandain 45
2-D cross-sections: segmentation
Diameter < 1.0 mm
derivatives with gaussian ofσ = 0.5 voxel
0.. =∇∇ IHII
16/01/2003 Grégoire Malandain 46
2-D cross-sections: segmentation
1250 1400
24
16/01/2003 Grégoire Malandain 47
Pre-processing?
• Isotropic diffusion
( ) )0(*)( IGtIIdivItI
t=⇒∇=∆=∂∂
• Anisotropic diffusion
( )( )IIII IIfIf
IIgdivtI
∇∇∇∇ −∆+=
∇∇=∂∂
21
|)(|
16/01/2003 Grégoire Malandain 48
Pre-processing?
25
16/01/2003 Grégoire Malandain 49
Tomographie à transmission et mouvement :application à l’angiographie rayon X cardiaque
16/01/2003 Grégoire Malandain 50
Motivations
• Contexte diagnostique et interventionnelléger :détection / quantification 3D de sténoses
• Contexte interventionnel lourd :modèle patient pour la chirurgiemini-invasive robotisée
26
16/01/2003 Grégoire Malandain 51
Angiographie cardiaquerotationnelle
16/01/2003 Grégoire Malandain 52
Problème• Voxels vi
• Pixels pj
• Rj vecteur des contributions des voxels au pixel pj
Rj.v = pjR.v = p
27
16/01/2003 Grégoire Malandain 53
Résolution
• Système linéaire de grande taille,sous-déterminé (+bruit de mesure)
• Méthodes (R, p) → v :• analytiques (FBP)• algébriques (ART, SIRT)• stochastiques (EM)• optimisation (CG)
16/01/2003 Grégoire Malandain 54
Tomographie statique
• Observation en transparence (rayons-X)• Plusieurs angles de vues• Objet statique
La matrice des contributions des voxels aux pixels est modifiée par le mouvement de l'objet observé
28
16/01/2003 Grégoire Malandain 55
Tomographie dynamique
• Pour chaque vue :• Pour chaque voxel :
• on déforme son centre• on projette le centre déformé• on l'attribue au pixel dans lequel la projection arrive
16/01/2003 Grégoire Malandain 56
Tomographie dynamique
• Calcul de la déformation :
• segmentation (analyse multi-échelle, intensité)• appariement• ajustement de faisceau• calcul de déformation 4D paramétrique
29
16/01/2003 Grégoire Malandain 57
Results - Matched views
16/01/2003 Grégoire Malandain 58
Segmentation
30
16/01/2003 Grégoire Malandain 59
Principle of binocular stereovision
16/01/2003 Grégoire Malandain 60
Principle of binocular stereovision
Binocular matching may be ambiguous
31
16/01/2003 Grégoire Malandain 61
Principle of trinocular stereovision
The third view often desambiguates the matching
16/01/2003 Grégoire Malandain 62
Principle of trinocular stereovision
The third view often desambiguates the matching
32
16/01/2003 Grégoire Malandain 63
Appariement / Ajustement
16/01/2003 Grégoire Malandain 64
Results - 3D Model Reprojection
33
16/01/2003 Grégoire Malandain 65
Déformation 4D
16/01/2003 Grégoire Malandain 66
Formation de l'image
• Intégrales des atténuations le long desrayons : I'(x) = -µ I(x)
• On néglige :• l’incohérence du faisceau• le rayonnement diffusé primaire (collimation)• le rayonnement diffusé secondaire (détecteur)• le post-traitement sur les images
∫=Rµ-e I(0) I(x) ))0(log())(log( IxI
R
−=∫ µ
34
16/01/2003 Grégoire Malandain 67
Résultats : séquence 4D (3D+t)