RT Journal Article
SR Electronic
T1 Comparison of 3 Methods of Automated Internal Carotid Segmentation in Human Brain PET Studies: Application to the Estimation of Arterial Input Function
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 461
OP 467
DO 10.2967/jnumed.108.059642
VO 50
IS 3
A1 Paolo Zanotti-Fregonara
A1 Renaud Maroy
A1 Claude Comtat
A1 Sebastien Jan
A1 Véronique Gaura
A1 Avner Bar-Hen
A1 Maria-Joao Ribeiro
A1 Régine Trébossen
YR 2009
UL http://jnm.snmjournals.org/content/50/3/461.abstract
AB Quantitative brain 18F-FDG PET studies often require the plasma time–activity curve (input function) for estimation of the cerebral metabolic rate of glucose (CMRglc). The gold standard for input function measurement is arterial blood sampling, which is invasive and time-consuming. Alternatively, input function can be estimated from dynamic images. This estimation often implies the use of manually placed regions of interest (ROIs) over cerebral vasculature, which is an operator-dependent and time-consuming task. The aim of our study was to compare 3 algorithms of image segmentation (local means analysis [LMA], soft-decision similar component analysis [SCA], and k-means) to automatically segment internal carotid arteries from dynamic 18F-FDG brain studies. Methods: The accuracy of automatic carotid segmentation algorithms was first tested using numeric phantoms of the human brain, by quantitatively assessing the overlap between the segmented carotids and the reference regions in the phantom. Then, the algorithm that yielded the best results was applied to data from 4 healthy volunteers, who underwent an 18F-FDG dynamic 3-dimensional PET brain study. Concordance between manual and automatic ROIs, both uncorrected and after partial-volume effect and spillover correction, was first assessed. Linear regression was then used to compare manual versus automatic CMRglc values obtained using Patlak analysis. CMRglc values obtained by arterial sampling were used as a reference. Results: In phantom studies, LMA was shown to be superior to the other segmentation algorithms. By visual inspection, volunteers' internal carotids segmented by LMA were anatomically relevant. No significant difference was found between ROI values obtained by manual and automatic segmentation, either uncorrected or corrected for partial-volume effect. Linear regression demonstrated excellent agreement between the manual and automatic image-derived CMRglc values (P < 0.0001), and both correlated well with the reference values obtained by plasma samples. Conclusion: The LMA segmentation algorithm allows accurate automatic delineation of internal carotids from dynamic PET brain studies. After correction for partial-volume effect, the main application would be the estimation of an image-derived input function.