TY - JOUR T1 - PET lung tumor delineation based on monotonicity and gradient features JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1444 LP - 1444 VL - 50 IS - supplement 2 AU - Cherry Ballangan AU - Xiuying Wang AU - Stefan Eberl AU - Michael Fulham AU - Dagan Feng Y1 - 2009/05/01 UR - http://jnm.snmjournals.org/content/50/supplement_2/1444.abstract N2 - 1444 Objectives FDG PET-CT is the best method to stage non-small cell lung cancer (NSCLC). The good boundary definition in CT can delineate isolated tumors. However, identification of tumor extension into the mediastinum and chest wall is problematic with CT and could be improved with PET-CT. SUV threshold and gradient based methods are prone to segmentation leaking when the tumor is located near other ‘hot spots’ or there is a low gradient. We propose a novel solution to delineate the primary tumor using SUV monotonicity and gradient magnitude. Methods Our method regards a tumor in a PET SUV image as a 3D monotonically decreasing function. Initially, the voxel with SUVmax is selected as the seed, and if its neighboring voxel vi has a decreasing value, vi is labeled as tumor and selected as the new seed. This procedure is repeated until vi’s normalized gradient magnitude GM(vi) reaches a threshold GMThresh. We define GMThresh=(1– SUV(vi)/SUVmax)3. This stopping criterion models the relationship between SUV(vi) and GM(vi) such that when SUV(vi) is low, GMThresh is high. Results The algorithm was evaluated in 14 NSCLC PET-CT volumes with pre-segmented tumors in CT. Tumors in PET delineated with different methods were validated on CT using volumetric overlap fraction (VOF) [1]. Our method’s average VOFs was comparable with the 50% SUVmax threshold and outperformed the other methods (Table 1) and no segmentation leaking occurred. Conclusions Our mathematical model of SUV gradient for lung tumor delineation solves the leaking problem in an automated delineation system. Research Support ARC and PolyU grants. ER -