Model based lesion segmentation in FDG-PET from raw data and clinical images

A review of recent works by our group in the segmentation and quantification of oncological lesions in 18-fluoro-deoxy-glucose (FDG) positron emission tomography (PET) images is given, stressing the underlying model assumption. In a first approach, a targeted reconstruction strategy was set in the framework of linear space variant (LSV) reconstruction from projections. Resolution recovery by ordered subset expectation maximization (OSEM) from raw emission data was based on the Poissonian description of errors and on a parametric model of PET scanner space variant blurring. The targeted strategy estimates and refines lesion basis functions while freezing the rest of background in a single one, thus providing an estimate of lesion borders and uptake. In a second approach lesions are segmented on standard clinical images. Reconstructed images lose the original Poissonian features and are well fitted by Gaussian mixture models (GMM). However, this popular clustering method required a specific adaptation to lesion segmentation: constraint of the warm background based on GMM modeling of healthy tissue; proximity priors in the voxel-wise classification to extract connected object and increase noise robustness. In both approaches, the spill-out of high uptake organs in the lesion area was removed by means of appropriate models relying on the CT anatomy. Results obtained on digital and physical phantoms and on clinical datasets will be recalled.