Optimisation and Validation of a Swarm Intelligence based Segmentation Algorithm for low Contrast Positron Emission Tomography

In the field of radiooncological research, individualised therapy is one of the hot topics at the moment. As a key aspect biologically-adapted therapy is discussed. Therapy adaption based on biological parameters may include tomographic imaging to determine biological properties of the tumour. One often invoked imaging modality is positron emission tomography (PET) using the tracer [18F]-fluoromisonidazole (FMISO) for hypoxia imaging. Hypoxia imaging is of interest, because hypoxic tumours are known to be radiorestistant. Even further, patients with hypoxic tumours have worse prognosis compared to patients with normoxic tumours. Thus, hypoxia imaging appears promising for radiotherapy treatment adaption. For example, volumetric analysis of FMISO PET could deliver additional hypoxia target volumes, which may be irradiated with higher radiation doses to improve the therapeutic effect. However, limited contrast between target volume and background in FMISO PET images interferes image analysis.Established methods for target volume delineation in PET do not allow determination of reliable contours in FMISO PET. To tackle this aspect, this thesis focusses on an earlier developed swarm intelligence based segmentation algorithm for FMISO PET and rather, its optimisation and validation in a clinically relevant setting. In this setting, clinical FMISO PET images were used which were acquired as part of a clinical trial performed at the Clinic and Policlinic for Radiation Therapy and Radiooncology of the University Hospital Carl Gustav Carus Dresden. The segmentation algorithm was applied to these imaging data sets and optimised using a cross-validation approach incorporating reference contours from experienced observers who outlined FMISO PET positive volumes manually. Afterwards, the performance of the algorithm and the properties of the resulting contours were studied in more detail. The algorithm was shown to deliver contours which were similar to manually-created contours to a degree like manually-created contours were similar to each other. Thus, the application of the algorithm in clinical research is recommended to eliminate inter-observer-variabilities. Finally, it was shown that repeated FMISO PET imaging before and shortly after the beginning of combined radiochemotherapy lead to manually-created contours with significantly higher variations than the variations of automatically-created contours using the proposed algorithm. Increased contour similarity in subsequently acquired imaging data highlights the observer-independence of the algorithm. While several observers outline different volumes, in identical data sets as well as in subsequent imaging data sets, the algorithm outlines more stable volumes in both cases. Thus, increased contour reproducibility is reached by automation of the delineation process by the proposed algorithm.