Segmentation of lymphoma tumor in PET images using cellular automata: A preliminary study

Abstract

Positron Emission Tomography imaging (PET) has today become a valuable tool in oncology. The accurate definition of the tumor volume on PET images is a critical step. State-of-the-art methods are based on adaptative thresholding and usually require user interaction. Their performances are hampered by the low contrast, low spatial resolution, and low signal to noise ratios of PET images. In this paper, we investigate an automated segmentation approach based on a cellular automata algorithm (CA). The method's results are evaluated against manual delineation on PET images obtained from 14 patients examinations obtained in clinical routine. Its performance is also compared to standard interactive PET segmentation algorithms (fixed or adaptive thresholding). Our method obtains an encouraging average Dice metric of 80.0%, a result comparable to the top methods. In case of small tumors, which are particularly difficult to segment, the method performs best among all of the state-of-the-art methods, both in terms of mean relative error volume (20.4%) and mean Dice metric (79.2%).