Effect of sample size for normal database on diagnostic performance of brain FDG PET for the detection of Alzheimer's disease using automated image analysis

Objective To investigate the relationship between the sample size for a normal database (NDB) and diagnostic performance of FDG PET using three-dimensional stereotactic surface projection for the detection of Alzheimer's disease. Methods We generated nine NDB sets consisting of 4, 6, 8,10, 20, 30, 40, 50 and 60 normal subjects. In order to assess the diagnostic performance using these NDBs to distinguish Alzheimer's disease patients from normal subjects, we recruited 52 patients with probable Alzheimer's disease (25 males, 27 females; mean age, 66.8 +/- 8.1 years) and 50 normal subjects (24 males, 26 females; mean age, 65.7 +/- 9.4 years). A receiver operating characteristic (ROC) analysis was performed for comparison of diagnostic accuracy among ND13 sets. Results Small N DBs (n <= 10) yielded poor quality of mean and SD images as compared with large NDBs (n >= 20). The ROC curves of the smaller group varied inconsistently, whereas those of the larger group were nearly superimposable. The area under the ROC curve (AUC) of the NDBs with sample size 6 (0.911) or 8 (0.929) was significantly smaller than that of the largest NDB (n = 60, 0.956). The AUCs of the larger group did not fall below 0.950, whereas AUCs of the smaller subgroup never exceeded 0.950. Conclusions Our data indicate that the sample size for an NDB affects the diagnostic performance of FDG PET using automated statistical approach, and that inclusion of at least 10 subjects is recommended, and 20 seems to be preferable for generating NDBs, although even a small NDB may provide clinically relevant results.