The development and testing of a digital PET phantom for the evaluation of tumor volume segmentation techniques

Abstract

Methods for accurate tumor volume segmentation of positron emission tomography (PET) images have been under investigation in recent years partly as a result of the increased use of PET in radiation treatment planning (RTP). None of the developed automated or semiautomated segmentation methods, however, has been shown reliable enough to be regarded as the standard. One reason for this is that there is no source of well characterized and reliable test data for evaluating such techniques. The authors have constructed a digital tumor phantom to address this need. The phantom was created using the Zubal phantom as input to the SimSET software used for PET simulations. Synthetic tumors were placed in the lung of the Zubal phantom to provide the targets for segmentation. The authors concentrated on the lung, since much of the interest to include PET in RTP is for nonsmall cell lung cancer. Several tests were performed on the phantom to ensure its close resemblance to clinical PET scans. The authors measured statistical quantities to compare image intensity distributions from regions-of-interest (ROIs) placed in the liver, the lungs, and tumors in phantom and clinical reconstructions. Using ROIs they also made measurements of autocorrelation functions to ensure the image texture is similar in clinical and phantom data. The authors also compared the intensity profile and appearance of real and simulated uniform activity spheres within uniform background. These measurements, along with visual comparisons of the phantom with clinical scans, indicate that the simulated phantom mimics reality quite well. Finally, they investigate and quantify the relationship between the threshold required to segment a tumor and the inhomogeneity of the tumor's image intensity distribution. The tests and various measurements performed in this study demonstrate how the phantom can offer a reliable way of testing and investigating tumor volume segmentation in PET.