Tumor segmentation for radiation therapy is one of the most important parts of treatment planning. Current treatment planning is based on an interpretation of the tumor's binary 3D map, which is created by automated or manual segmentation of CT scans. The latter is used more frequently in clinical practice. The subjective nature of manual segmentation can lead to great observer variability and segmentation uncertainty, which is the greatest at the edge of target volumes. This uncertainty can lead to a inappropriately distributed radiotherapy dose and poor repeatability of treatment planning.
In this master thesis, we have developed a new approach to defining a tumor volume that could reduce the effect of uncertainty and could potentially lead to increased repeatability of treatment. We have developed a new method for specifying target volumes based on a PET and a CT scan by creating a probabilistic tumor likelihood map in the head and neck area. It contains the high and low likelihood tumor maps and target volumes uncertainty assessment.
Depending on the radiotracer uptake in each voxel of a PET scan, the model first creates a high likelihood map of disease presence. Depending on the distance and surrounding anatomy, model calculates the likelihood of microscopic infiltration. In this part, the map of high likelihood is expanded to a map of low likelihood for tumor presence.
The method also includes the evaluation of tumor likelihood map uncertainty, which has two main contributions - imaging uncertainties and uncertainty of microscopic tumor infiltration. Imaging uncertainties are assessed based on known uncertainties of a PET scan, whereas the uncertainty of microscopic infiltration is calculated based on modeling results at different parameter values of microscopic tumor infiltration.
Tumor likelihood maps were compared to clinical segmentation -- the binary maps of tumor presence. Individual axial slices show that in some areas the likelihood and the binary map of tumor presence overlap for both analyzed patients. In other parts, area of non-zero likelihood for tumor presence was smaller than the binary map of clinical segmentation. Uncertainty of the model and tumor presence likelihood map is a combination of microscopic infiltration uncertainties and image uncertainties. The results show that the latter was the dominant factor for the analyzed cases, with its contribution being approximately three times greater than the uncertainty of microscopic tumor infiltration. Use of the probability based approach of defining tumor presence and the evaluation of uncertainty based on the two patients used in the analysis serves as a good proof of concept. Only an expanded study with a larger population size and appropriately long follow up would be needed for better repeatability of specifying target volumes and treatment planning in radiotherapy.
|
