TU‐C‐211‐12: Performance of a Novel Region Growing Segmentation Method for Nuclear Medicine

Abstract

Purpose: To investigate the performance improvement of a novel region growing (SRG) and “segmentation‐by‐consensus” (SBC) technique for delineation of lesions in FDG‐PET. Further, the improvement in lesion segmentation using time‐of‐flight (TOF) PET was investigated. Methods: The NEMA image quality phantom was filled to give 3:1 and 7:1 contrast ratios between spheres (0.44 – 27.13 cc) and background. Contrast added to the sphere volumes and a hi‐resolution CT allowed segmentation of the true activity distribution. Ten minute list‐mode scans were acquired on a GE Discovery 690 PET/CT system. Images were reconstructed representing 1, 2 and 5 minute acquisitions using TOF and NTF methods. Four commonly used segmentation methods, 40%max 50%max, K‐means, fuzzy C‐means, and the SRG technique, were used to segment the six spheres. The SBC approach creates volumes from voxels segmented by a majority of the techniques. Absolute volume differences and volume overlap were metrics for evaluating each algorithmˈs performance. Results: All techniques performed poorly for the small (0.44 & 1.08 cc) spheres with mean volume errors between 10–30 cc (3:1 contrast), and 5.5–12 cc (7:1). SRG consistently had lower average volume errors (4.25–6.30 cc) but 40%max had the highest absolute volume overlap for the largest sphere (0.7795– 0.8090). However, SRG dramatically outperformed 40%max for the 1.08 cc sphere (0.4338–0.4952 vs. 0.0124–0.0173). For the high (7:1) contrast, k‐ means consistently gave lower average volume errors (0.09–6.95 cc) and higher average volume overlap (0.4741–0.5765) than SRG (8.57–10.15 cc and 0.3543–0.3967). No noticeable difference with TOF was observed. SBC gave results within the above ranges for both 3:1 and 7:1 with less variation and was close to the highest performing technique in each scenario. Conclusions: TOF offers comparable results to NTF PET for the volumes segmented. The SRG technique provided lower mean volume errors and good volume overlap in low contrast situations.