Multi-tracer analysis for patient’s following using multi-observation statistical image fusion : a feasibility study S. David1, M. Hatt1, P. Fernandez2, M. Allard2, O. Barrett2, D. Visvikis1 1. LaTIM, INSERM U650, Brest, France 2. Department of nuclear medicine Hospital Pellegrin – CHU Bordeaux Outline Context of oncology • Positron Emission Tomography (PET) • External radiotherapy Use of PET in clinical application • Multi-tracer analysis for dose-painting • Patient monitoring in PET Multi-observation fusion • Multiband segmentation of a spectroscopic line data cube • Developed method • Preliminary results Further work Context of oncology Cancer • In 2002 : 11 millions new cases and 7 millions deaths • Foresee in 2030 : 11 millions deaths Diagnosis • Computed tomography (CT) • Magnetic resonance imaging (MRI) • Emission imaging (PET, SPECT) Since 2000 : PET / CT combined Gold standard for the diagnosis Treatment • Surgery • Chemotherapy • Radiotherapy Usually combined Focus on the PET imaging and its application : • Radiotherapy planning • Therapy response assessment Context and motivations Positron Emission Tomography (PET) : • Functional imaging : visualization of physiological processes • Mainly used for cancer diagnosis and staging Principle : • Injection of a radiotracer o Biological tracer targets the tumor o Radionuclide : β- emitter • Detection of the 2 γ rays • Image reconstruction by tomography Drawbacks of the PET imaging : • Blur (spatial resolution) • High noise (acquisition variability) • Low resolution ( >5 mm) Context and motivations PET/CT imaging : • Since 2000, acquisition of PET and CT in the same bed • Gold standard for the diagnosis CT image PET image PET / CT fusion + Combination of anatomical and functional information + Allow to anatomically locate the tracer uptake in the PET image - Registration of the CT and PET scans - Difference in the image resolution ( CT <1mm, PET >5 mm) PET / CT used in the radiotherapy planning Context and motivations Principle of radiotherapy : • Use the ionizing radiation to kill the malignant cells External radiotherapy (most widely used) : • Photon or electron beams produced by a linear accelerator • Shape of the beam defined by the collimator Definition of the target volumes : • Gross tumor volume (GTV) o Defined with conventional imaging modalities • Clinical target volume (CTV) > GTV o Volume computed considering the anatomical information • Planning target volume (PTV) > CTV o Take into account the physiological variations Use of PET in clinical application Biological image-guided dose escalation • Development of Intensity-modulated radiation therapy (IMRT) • Administration of a non-uniform dose o Adapt the treatment to the patient o Reduce the irradiation of organ at risk (OAR) and surrounding healthy tissues PET / CT : Target volume definition IMRT planning Improvement of the tumor volume definition with the multi-tracer analysis Multi-tracer analysis • 18F-FDG : Measure the glucose consumption (tumor highly glucose consumer) o Radiotracer the most widely use in PET o Not tumor specific (other physiological processes need glucoses) o Uptake in inflammatory tissues Use of PET in clinical application Development of specific radio-marker measuring ≠ tumor features : • FMISO : measure of the tumor hypoxia (lack of oxygen) o Hypoxia induces resilience to the radiotherapy Hypoxic tumors require a dose boosting • 18F-FLT : measure of the tumor proliferation oTumor specific radio-marker : ≠ FDG o Avoid inflammatory tissues o Lower uptake in the tumor than FDG FDG coronal PET scan • Merging all features measured by the tracers • Each radiotracer is measuring a specific biologic process o Scans are quantitatively not comparable Y. Yamamoto et al, European Journal of Nuclear Medicine and Molecular Imaging , 2008 FLT coronal PET scan Use of PET in clinical application Patient follow-up with PET : • Patients underwent chemo / radio-therapy o Early assessment of the response to the treatment • PET acquisition during the course of the therapy o Estimation of prognosis Tumors PET1 pre-treatment PET2 post-treatment o Adapt the therapy to the response o Avoid toxic and costly cares for non-responding patients Goal of our work Multi-sensor observation of an objet • Many clinical data available with the multi-modality imaging Goal : • Fusing all the scans obtained : o with the ≠ radiotracers Should improve the tumor volume definition o at different time of the treatment A more accurate assessment of therapy response Approach : • With the statistical segmentation framework • Model the information of multi-tracer and/or follow-up scans Analogy with the astronomical framework Multiband segmentation of a spectroscopic line data cube Segmentation process based on Bayesian inference • Observation with radio interferometers o 3D data cubes (astronomical coordinates and frequency third axis) Reduction of the dimension (42 channels) of the cube • EM algorithm on Gaussian mixture model o fit the spectrum : assessment of mean and variance • Choice of the 6 most pertinent Gaussian o K-means • Computation of weight associated to each class o Levenberg-Marquardt algorithm Maps of the weights of the 6 Gaussian with the NGC 4254 Cube F. Flitti et al, Multiband segmentation of a spectroscopic line data cube : application to the H1 data cube of the spiral galaxy NGC 4254 Multiband segmentation of a spectroscopic line data cube Bayesian classifier • Hierarchical Markovian Model o Models the spatial dependencies between neighbors Steps of the segmentation process • Initialization : K-mean algorithm • Parameter estimation step o Unsupervised with the Iterative Conditional Estimation (ICE) • Segmentation step o Criteria of maximum a posteriori (MAP) Results on the NGC 4254 Cube • Creation of a label map F. Flitti et al, Multiband segmentation of a spectroscopic line data cube : application to the H1 data cube of the spiral galaxy NGC 4254 Multi-observation method Statistical image segmentation : • Estimation of X ( xt )tT (hidden) with Y ( yt )tT (measurements) • No determinist link between X and Y o Probabilistic approach (Bayesian inference) p ( x, y ) p ( y | x ) p ( x ) Global prior model (spatial or contextual) Local Observation model (noise) Iterative estimation of the parameters Segmentation Markovian model (field, quadtree…) Blind, contextual, adaptive… Gaussian, generalized gaussian deterministic (EM) stochastic (SEM) hybrid (ICE) MAP, MPM criteria Multi-observation method • X : labels field • Y : measurements field Measures Y time PET FDG … PET MISO ... … … … … PET Tracer N … … … … ... ... Label X Multi-tracer analysis Estimation of the (X,Y) distribution • Mixture defined by θ=(α,β) o αi : priors of X o βi=(mi,Гi) : distribution of Y conditional to X Label X Patient follow-up … … … … Multi-observation method : preliminary work Fusion process : Initialization : • K-means • Fuzzy K-means Estimation of the (X,Y) distribution • Maximization of the log-likelihood • Implementation of EM and SEM algorithm o Blind and adaptive (AEM,ASEM) version Decision step • MAP criteria o Creation of a segmented map Test on data set : • Synthetic images • Simulated tumors Preliminary results Synthetic images : Map of labels X • Map of labels X o 2 or 3 labels per image • Measurements Y o Mean discrimination (MD) o Variance discrimination (VD) • Fusion of measurement with the same map of labels o Number of labels in segmentation map = Number of labels per image Segmentation results : Measurement Y Segmentation map • N = 2 spectral bands • K = 2 labels • Random initialization MD VD MD VD AEM segmentation ASEM segmentation • N = 2 spectral bands • K = 3 labels • Random initialization AEM segmentation ASEM segmentation Preliminary results Synthetic images : • 2 or 3 labels per image • Fusion of measurement with the different map of labels o Additional labels in the segmented map Segmentation results : Measurement Y Segmentation map • N = 2 spectral bands • K = 2 and 3 labels • 4 classes in segmentation map • Fuzzy K-means initialization AEM segmentation ASEM segmentation AEM segmentation ASEM segmentation • N = 2 spectral bands • K = 3 labels per scan • 5 classes in segmentation map • Fuzzy K-means initialization Preliminary results Simulated tumors : The ground truth of each tumor scan will be different The segmentation process should identify the new labels • Each band : a tracer scan • Label in the scan : uptake of the tracer 3 cases : • N = 2 tracer scans • K = 2 levels of tracer uptake per scan • Fuzzy-K mean initialization • N = 3 tracer scans • K = 2 levels of tracer uptake per scan • Fuzzy-K mean initialization • N = 2 tracer scans • K = 3 levels of tracer uptake per scan • Fuzzy-K mean initialization Preliminary results Segmentation results : • N images = 2 , K tracer uptake =2 AEM segmentation ASEM segmentation • N images = 3 , K tracer uptake =2 AEM segmentation ASEM segmentation AEM segmentation ASEM segmentation • N images = 2 , K tracer uptake =3 Preliminary results Fusion of the synthetic images • In the different situations (N Bands, K classes) o Supervised, semi-supervised and unsupervised segmentation o Satisfactory classification Fusion of simulated tumors • Segmentation error depends on : o The fuzzy K-means initialization o Noise level in the scans o The number of classes in the label map Limitations : • Segmentation not totally unsupervised : o Number of labels has to be defined by the user • Fusion of few spectral bands Further work • Fusion process with more data o Other tracers images and / or follow-up scans • Segmentation totally unsupervised o Estimation of the classes number in the label map • Test the method on simulated data with GATE o More realistic simulated tumors o Computation of classification error • Application of our method in a radiotherapy planning station Thank you for your attention