Abstract
1301
Objectives Various radiotracers measuring specific tumour metabolism can be used in positron emission tomography (PET) imaging in order to assess therapy response. Quite frequently it may be useful to establish an overall analysis of the tumour response considering all the available information provided by the different radiotracers. However, managing several different scans may increase the difficulty in the interpretation, thus hampering the benefit of having complementary images.
Methods The statistical approach assumes the data can be modeled by a mixture distribution of random fields (X,Y). The multi-tracer scans, acquired at different treatment stages, can be considered as a multi-dimensional measurement Y of the hidden labels X, representing the different activity concentrations. The distribution (X,Y) is estimated using the Stochastic Expectation Maximization (SEM) algorithm with adaptive local priors (ASEM). Lastly, a segmentation and fusion map of the labels X is created, reflecting each tracer distribution. The fusion process has been tested on two kinds of datasets, synthetic images and simulated tumors, considering different numbers of classes and tumour shapes. The fusion process evaluation is obtained by computing volume errors of the segmented map.
Results For the synthetic images, the segmentation process led to satisfactory results with a mean volume error of 6.6 ± 11.8% across the different cases. Similarly, for the simulated tumours, the mean volume error assessed for the different configurations was 8 ± 4.9%. The volume error of the segmentation process depends on the tumour shape, the number of classes of each scan and the number of images included in the fusion.
Conclusions Further work will investigate applying this segmentation/fusion process on more realistic datasets corresponding to additional tracers or follow-up scans. Validation on simulated and clinical datasets in the context of therapy assessment in oncology applications will also be carried out using this new methodology