Download presentation
Presentation is loading. Please wait.
1
From nebulae segmentation in astronomical imaging to tumor delineation in 18F-FDG PET imaging: how can one serve the other? M. Hatt 1, C. Collet 2, F. Salzenstein 3, C. Roux 1, D. Visvikis 1 Speaker: S. David 1 1. LaTIM, INSERM U650, Brest, France 2. LSIIT, CNRS - UMR 7005, Strasbourg, France 3. INESS, CNRS - UMR 7163, Strasbourg, France
2
Context and objective Cancer Oncology Gold standard for diagnosis Other applications of interest: Radiotherapy planning Prognosis, therapy assessment PET/CT multimodality imaging Quantification active biological volume uptake measurement radiotherapy target definition Requires the definition of a volume of interest Computed tomography (CT) Positron Emission Tomography (PET) Source of imageX-rayPositron emitter ( 18 F) NatureAnatomic: tissues and bones density Functional : accumulation of radioactive tracer Resolution< 1 mm> 5 mm Imaging for oncology
3
Context and objective Problems of PET images 3 Noise (acquisition variability) Blur (spatial resolution) Voxels size (grid spatial sampling) uptake heterogeneities within the tumor
4
Methodologies Existing solutions Manual definition of regions of interest in the background Parameters optimization for each scanner Assume tumors are homogeneous spheres : Threshold-based methodologies [1-3] [1] J. A. van Dalen et al, Nuclear Medicine Communications, 2007 [2] U. Nestle et al, Journal of Nuclear Medicine, 2005 [3] J.F. Daisne et al, Radiotherapy Oncology, 2003 Require a lot of a priori information and are system and user dependent But tumors are often of complex shapes and heterogeneous !
5
PET images share several characteristics with some astronomical images Why looking at astronomical images processing for solutions ? The segmentation/classification field is more mature for astronomy than PET Methodologies Astronomical images segmentation
6
Nebulae vs PET tumor ? Methodologies Astronomical images segmentation
7
Nebulae vs PET tumor ? CharacteristicNebulae imagePET tumor image Dimensions2D, multi/hyper spectral3D, mono spectral DefinitionLarge (~512x512)Small (~30x30x30) Encoding32b real16b/32b real Fuzzyyes Noisyyes Band 1Band 2 Band 3 Slice n+1 Slice n Slice n-1 Use of statistical image processing to deal with the noise, combined with fuzzy modeling to deal with blur Methodologies Astronomical images segmentation
8
Methodology : statistical + fuzzy Probabilistic / statistical part models the uncertainty of classification Fuzzy part models the imprecision of acquired data Combining both to model astronomical or PET images characteristics : Discrete Dirac measure on class c Standard (“hard”) statistical modelling Ground-truth : Continuous Lesbegue measure on Fuzzy modelling [1] [2] [1] H. Caillol et al, IEEE Transactions on Geoscience Remote Sensing, 1993 [2] F. Salzenstein and W. Pieczynski, CVGIP : Graphical Models and Image Processing, 1997 Methodologies
9
Methodology: fuzzy Markov chains Markov assumption: …… Transition probabilities Initial probabilities Use of the Hilbert-Peano path to transform 2D image into 1D chain Observation vector in [0,1] Methodologies
10
Result on Nebulae Fuzzy Hidden Markov Chains (FHMC) multispectral segmentation F. Salzenstein, C. Collet, S. Lecam, M. Hatt, Pattern Recognition Letters, 2007 Methodologies
11
Apply to PET ? 3D PET tumor Iterative stochastic estimation (SEM) 1D chain with discrete values {0,1,F 1,F 2 } Segmentation (MPM) 1D chain with real values Hilbert-Peano 3D Inverse Hilbert- Peano 3D Segmentation map (2 fuzzy levels) Extended Hilbert-Peano path to transform 3D image into 1D M. Hatt et al, Physics in Medicine and Biology, 2007;52(12):3467-3491 Methodologies
12
Problem ! 3D Hilbert-Peano path to transform 3D image into 1D disrupts spatial correlation : Neighbors voxels in the image may be far from each other in the chain Size of tumors with respect to object and size of voxels leads to large errors for small tumors ! M. Hatt et al, Physics in Medicine and Biology, 2007; 52(12):3467-3491 Methodologies
13
Solution: locally adaptive method 3D PET tumor Segmentation map FHMC M. Hatt et al, IEEE Transactions on Medical Imaging, 2009;28(6):881-893 Iterative stochastic estimation (SEM) Segmentation Markovian model replaced by sliding estimation cube to compute probabilities for each voxel regarding its neighbors : FLAB (Fuzzy Locally Adaptive Bayesian) method Methodologies FLAB
14
1 2 3 M. Hatt et al, International Journal of Radiation Oncology Biology Physics, in press, 2009 Modeling fuzzy transitions between pairs of hard classes to deal with heterogeneities 2 hard classes and 1 fuzzy transition 1 0 Methodologies 3 hard classes and 3 different fuzzy transitions
15
Simple phantom validation Results Phantom acquisitions with spheres : 37 to 10 mm in diameter PhantomComputed tomography image (truth) 18 F Positron Emission Tomography image AxialCoronalSagital
16
Results FHMC vs FLAB M. Hatt et al, IEEE Transactions on Medical Imaging, 2009;28(6):881-893
17
Multiple scanners robustness validation 4 different scanner models and various acquisitions parameters (contrast, noise, reconstruction algorithms, size of voxels…) Philips GeminiGE Discovery LS OSEM Siemens Biograph RAMLA 3D Philips Gemini TF TF MLEM A B 1 211 2 1 2 A = 4:1 or 5:1, B = 8:1 or 10:1 1 = 2x2 mm, 2 = 4x4 or 5x5 mm 37 mm28 mm22 mm17 mm13 mm M. Hatt et al, Society of Nuclear Medicine annual meeting, Toronto, Canada, 2009 Results
18
Real Simulated Small homogeneous Large heterogeneous RealSimulated 20 tumors (NSCLC, H&N, Liver) maximum diameter from 12 to 82 mm Heterogeneities: from none to high Shapes: from almost spherical to complex Simulated with Monte Carlo GATE (Geant4 Application for Tomography Emission) M. Hatt et al, International Journal of Radiation Oncology Biology Physics, in press, 2009 Results Accuracy validation on simulated data
19
FLAB Ground-truth Fixed threshold Classif. error: 6%> 100% Simulated PET Adaptive threshold Classification errors Grey region 4% Black region 2% Volume error -62% Volume error +37% Segmentation Adaptive threshold FLAB Fixed threshold Ground-truth Simulated PET 14% M. Hatt et al, International Journal of Radiation Oncology Biology Physics, in press, 2009 Results Accuracy validation on simulated data
20
Patients with histology accuracy validation 18 tumors (NSCLC) with histology study [1] maximum diameter from 15 to 90 mm (mean 44, SD 21) Heterogeneity : none to high Shapes : from almost spherical to complex CT PET [1] A. van Baardwijk et al, International Journal of Radiation Oncology Biology Physics, 2007 Results
21
Patient with NSCLC FLAB Adaptive thresholdFixed threshold (42%) M. Hatt et al, International Journal of Radiation Oncology Biology Physics, in press, 2009 Results Patients with histology accuracy validation
22
Conclusions and work in progress Studies are ongoing to further investigate the clinical impact of the proposed methodology in radiotherapy or patient prognosis and therapy assessment This work is a good example of know-how transfer from astronomical to medical imaging Once adapted to PET data (2D->3D, spatial modeling), statistical and fuzzy segmentation developed for astronomical imaging performed admirably well for tumor delineation
23
Thank you for your attention
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.