The technique of anatomic standardization and comparison with normal templates is increasingly used in clinical brain SPECT practice and allows automated, operator-independent volume-of-interest (VOI) or voxel-based analysis of whole-brain data. In 2 distinct clinical populations with severe traumatic brain injury and cognitive impairment, this study compared 3 widely available approaches that use normal templates to evaluate SPECT brain perfusion deficits. Methods: In total, 74 subjects were studied. These included 14 patients with severe, traumatic brain injury (group 1; 10 males, 4 females; mean age ± SD, 27.6 ± 8.2 y) and 15 patients with cognitive impairment (group 2; 7 males, 8 females; mean age, 75.8 ± 8.6 y). These data were compared with those from, respectively, 25 and 20 age- and sex-adjusted healthy volunteers. All data were analyzed in 4 ways. Three semiquantitative statistical algorithms were used: statistical parametric mapping (SPM) using SPM99, brain registration and analysis of SPECT studies (BRASS) using a voxelwise region-growing technique, and a predefined VOI approach. These results were compared with visual analysis based on consensus reading by 3 experienced nuclear medicine physicians. Receiver operating characteristic (ROC) analysis was performed at various statistical cutoffs. Moreover, as a measure of regional agreement, relative regional agreement between methods was assessed. Results: In both study groups, BRASS voxel-based analysis was most accurate, as defined by the area under the ROC curve (0.97 for group 1 and 0.96 for group 2). VOI assessment was slightly more accurate than visual consensus analysis, whereas SPM showed, overall, a lower area under the ROC curve. SPM analysis was also significantly less sensitive at thresholds corresponding to low false-positive fractions. Regional analysis showed 83%-92% agreement between all methods. Conclusion: Under clinical conditions, classification of brain SPECT studies can greatly be aided by anatomic standardization techniques and reference to normal data. Under the investigated circumstances, SPM was found to have a lower sensitivity than VOI or voxelwise region-growing techniques, especially at low false-positive fractions.
