3D bayesian segmentation with poisson mixtures for brain tumoral and healthy tissue analysis in dual phase 18-FDG PET

Objectives: In the setting of a clinical protocol using brain dual phase 18-FDG PET we developed a segmentation tool to draw 3D Regions Of Interest (ROI) allowing to compute 3D quantifications on pathological and healthy tissues. Methods: The segmentation tool we implemented and tested is based on the modeling of the levels in the image with probabilistic mixtures, the estimation of the parameters of the mixture by means of a Stochastic Expectation Maximization (SEM) algorithm and the segmentation of the 3D PET volume with an Iterative Conditional Mode (ICM) algorithm. We compared Gaussian mixture and Poisson mixture based segmentations with threshold-based segmentations on simulated volumes. The best segmentation tool has then been used to segment real data and to compute descriptive statistics on white matter, gray matter, and tumors. 3D quantifications has been compared to a 6-level semi-quantitative visual analysis. Results: Poisson mixture based segmentation provided better results on simulated volumes than Gaussian mixture based segmentation does. So we used Poisson mixture based segmentation on real data. However, pathological or healthy tissues are still covered by multiple regions, each region being associated with one of the component of the mixture. The physician still has to merge some regions in order to obtain a final ROI exploitable for computing descriptive statistics. The diagnoses using descriptive statistics obtained from ROI were tallied with the visual diagnoses. Conclusions: We provided an accurate 3D implementation of SEM and ICM algorithms based on probabilistic mixture and Potts model in order to segment 18-FDG PET images