Validation of Automated PET Segmentation Methods Based on Connected Components for Myocardium

Automated segmentation of myocardium is essential as it allows reproducible and fast image analysis. In this study, three methods for automated myocardial segmentation were validated. The methods were based on the connected components (CC) derived from Positron Emission Tomography (PET) images. PET images were acquired on 22 patients with [
¹⁸
F]- Fluorodeoxyglucose ([
¹⁸
F]-FDG). The methods were compared against manual segmentation in terms of Dice indices as well as volume and center-of-mass (COM) differences. The developed methods are called the largest of CCs (LCC), number of CCs (NCC) and supervised CCs (SCC). LCC method is based on deriving the largest CCs, NCC deriving the number of CCs and SCC considers CCs that contain a voxel that locates in the myocardium. Dice indices for these methods (mean ± SD) were 0.70 ± 0.35, 0.54 ± 0.45 and 0.84 ± 0.17 for LCC, NCC and SCC, respectively. Volume differences were (mean ± SD) 0.73 dL ± 0.65 dL, 4.78 dL ± 5.77 dL and 0.58 dL ± 0.63 dL and COM differences were (mean ± SD) 16.9 mm ± 36.7 mm, 46.3 mm ± 58.1 mm and 2.47 mm ± 2.54 mm for LCC, NCC and SCC, respectively. The SCC method had the highest Dice indices with lower volume and COM differences compared to NCC and LCC.