Evaluation of linear registration algorithms for brain SPECT and the errors due to hypoperfusion lesions

• Published in: Medical physics (Lancaster) Vol. 28; no. 8; pp. 1660 - 1668
• Main Authors: Radau, Perry E., Slomka, Piotr J., Julin, Per, Svensson, Leif, Wahlund, Lars-Olof
• Format: Journal Article
• Language: English
• Published: United States American Association of Physicists in Medicine 01-08-2001
• Subjects: Aged; Aged, 80 and over; Algorithms; Alzheimer Disease - diagnosis; Alzheimer's disease; brain; Brain - pathology; Case-Control Studies; Computed tomography; Error analysis; Female; Humans; Image analysis; image registration; Least-Squares Analysis; Linear Models; Male; measurement errors; medical image processing; Medical imaging; Medicin och hälsovetenskap; Middle Aged; Models, Statistical; mutual information; Neuroscience; Perfusion; Physicists; registration; Reproducibility of Results; Retrospective Studies; segmentation; single photon emission computed tomography; Single photon emission computed tomography (SPECT); Software; Telencephalon - pathology; Time Factors; Tomography, Emission-Computed, Single-Photon - methods; segmentation; registration; mutual information; Alzheimer’s disease; single-photon emission computed tomography
• ISSN: 0094-2405; 2473-4209
• DOI: 10.1118/1.1388894

The semiquantitative analysis of perfusion single-photon emission computed tomography (SPECT) images requires a reproducible, objective method. Automated spatial standardization (registration) of images is a prerequisite to this goal. A source of registration error is the presence of hypoperfusion defects, which was evaluated in this study with simulated lesions. The brain perfusion images measured by
99m
Tc-HMPAO SPECT
from 21 patients with probable Alzheimer’s disease and 35 control subjects were retrospectively analyzed. An automatic segmentation method was developed to remove external activity. Three registration methods, robust least squares, normalized mutual information (NMI), and count difference were implemented and the effects of simulated defects were compared. The tested registration methods required segmentation of the cerebrum from external activity, and the automatic and manual methods differed by a three-dimensional displacement of
1.4±1.1 
mm
.
NMI registration proved to be least adversely effected by simulated defects with 3 mm average displacement caused by severe defects. The error in quantifying the patient-template parietal ratio due to misregistration was 2.0% for large defects (70% hypoperfusion) and 0.5% for smaller defects (85% hypoperfusion).