This study aimed to assess whether the event of SQA and the evolution of digital PET imaging requires the control databases to evolve concomitantly. Diagnostic performances observed at the group and individual level show that the diagnostic accuracy of SQA on digital controls is improved compared to SQA on conventional controls, particularly as it relates to the detection sensitivity of AD. This observation is an argument which supports the development of digital control databases for SQA of brain 18F-FDG PET images for clinical practice.
Digital PET technology is associated with improvements in image quality, specifically better spatial resolution and signal-to-noise ratios compared to conventional PET cameras (2). These distinct image qualities lead to problematic head-to-head comparisons between digital and conventional PET images as reflected in our study by the relatively poor diagnostic performance obtained for SQA using conventional controls (accuracy at the individual level of only 63%). However, all currently implemented control databases in dedicated software for automated SQA in clinical practice still rely on conventional PET control images (9, 13). Of course, it is now possible to implement local databases in the majority of these types of software, but establishing control databases acquired with digital PET technology remains an extensive undertaking, particularly because it involves a relative recently implemented technology.
Our present study shows that implementing a control database acquired with digital PET technology yields an increase in the detection sensitivity of AD patients, not only at the group level (+ 34 cm3 of detected hypometabolism volume, Table 2) but also at the individual level (sensitivity of detection increased from 37 to 89%, Table 4). A high detection sensitivity is primordial in the diagnosis of AD since 18F-FDG PET is a biomarker of neurodegeneration, which contributes to the ATN classification, the N biomarker being directly associated with cognitive impairment in patients suspected to have neurodegenerative diseases (14).
All results in the current study were initially obtained using a fully automated analysis, which supports the objective nature of our observations in both the group and individual level analyses. This original fully automated methodology, which necessitated an adaptation of the levels of significance to detect anomalies, was exclusively based on the SPM software. From a clinical standpoint and at the individual level, this fully automated analysis was nevertheless consistent with the visual analysis, using a methodology that is very similar to that applied in previously published SQA studies (6, 12). By using this visual analysis, the diagnostic performance of SQA with the digital controls observed in our study (sensitivity and accuracy of respectively 89 and 78%) was within the range of previously reported results (62.3–96% for sensitivity and 70-97.5% for accuracy), but tended towards a lower specificity (64% for a reported range of 84–99%) (6, 8–12, 21–23). This tendency to a lower specificity observed in our study is probably the result of a compromise with regards to the excellent sensitivity reported. This observation was similarly detected in previously reported results. Perani et al (6) reported high sensitivity but lower specificity for visual SQA of suspected neurodegenerative disease patients (respectively 96% and 84%). A sensitivity of 82% and a lower specificity of 75% were also reported by Lehman et al (12) when comparing MCI and AD patients to controls with a SQA.
SQA at the individual level in our visual analysis was performed using an intensity normalisation based on the sensitive-motor cortex as advocated by Yakushev et al (17). This is different to previously published SQA which performed an intensity normalisation based on the proportional scaling (6). In any case, the same visual analysis performed with PET images normalised to the proportional scaling yielded similar results with regards to accuracy, sensitivity and specificity with differences of 78% vs. 73%, 93% vs. 63%, and 59% vs. 86% observed for SQA using digital and conventional controls respectively (results not shown, p < 0.01 for the comparison of both control databases).
The low sensitivity observed with the SQA using conventional controls (37%) could be related to the unusual post-filter applied to conventional PET images (Gaussian kernel of 4 mm FWHM, predominantly adapted to the digital PET technology). However, when smoothing conventional PET images with a more adapted isotropic 3D Gaussian kernel of 8 mm FWHM, the accuracy of SQA using conventional controls remained unchanged (65%), with an increased sensitivity (93%), but at the expense of a larger decrease in specificity (32%).
The main limitation of our study results from the fact that controls included in the conventional and digital control databases were different individuals. From an ethical perspective, it remains problematic to establish control databases acquired in parallel with both the conventional and digital PET systems. It should however be noted that controls included in our conventional and digital databases did not exhibit any differences in age, sex, MMSE and educational level when compared to each other, or when compared to the AD and H groups. In addition, these two, distinct conventional and digital control databases are representative samples from current daily clinical practice. The main objective of the current study was to assess whether there is indeed a requirement to establish digital control databases when acquisitions are performed with the new digital PET system. A secondary issue that may be addressed is that the sample size of conventional and digital control databases is rather small (n = 19 and 20). This number of controls is nevertheless known to be sufficient to accurately perform group analyses with SPM (24) .
Overall, in light of recent digital PET technology developments and considering that SQA is now clearly recommended for brain 18F-FDG PET image analysis, there is an urgent need to establish digital PET control databases for SQA of brain 18F-FDG PET images. This would be particularly helpful for improving the sensitivity required to detect AD patients.