PET functional volume segmentation: a robustness study

Current state of the art algorithms for functional volume segmentation in PET images for diagnosis, patients follow-up or radiotherapy treatment planning consist of adaptive threshold approaches. We have developed an unsupervised Bayesian segmentation algorithm for tumors in PET, namely the FLAB (for Fuzzy Locally Adaptive Bayesian) algorithm, that was previously validated on simulated images and then successfully extended and applied to inhomogeneous and non spherical real clinical images of lung tumors. In this study, we investigate the robustness of this approach in comparison to fixed threshold based approaches at 42% and 50% of the maximum value, as well as another automatic segmentation algorithm (the fuzzy C-means clustering). For this investigation, a series of IEC phantom acquisitions were performed on different PET/CT scanners (Philips Gemini, Siemens Biograph and GE Discovery LS) and reconstruction algorithms (RAMLA, OSEM). Various acquisition parameters were considered in each case, like the size of the voxels in the reconstructed image, the contrast ratio and the acquisition duration (hence the intensity of the noise). The purpose of this study was to study the robustness of each approach with respect to scanner model, reconstruction algorithm, and various acquisition parameters. For all the scanner types and reconstruction algorithms considered in this preliminary study the FLAB algorithm demonstrated globally higher accuracy in delineation of the spheres. In addition, the FLAB results showed lower variability with respect to changes in the acquisition parameters and/or the scanner model and associated reconstruction algorithm, therefore demonstrating its higher robustness compared to the other semi-automatic or automatic approaches considered. Future studies will investigate the results of the same approaches with GE DST and Time of flight scanners, as well as the inclusion of the adaptive thresholding approaches.