We evaluated the clinical performance of a recently developed head motion correction algorithm for amyloid brain PET images. Image quality and interobserver agreement significantly improved after motion correction. The difference in UR between the conventional and motion-corrected PET images was significantly greater in the group with head motion than in the group without head motion.

Unlike the lung, heart, or bowel, the brain exhibits negligible autonomous movement. However, amyloid brain PET requires acquisition times of typically 10–20 min, and patients undergoing amyloid PET are suspected to have dementia and are often unable to tolerate the study. Therefore, there is a high likelihood that subtle to large movements occur during PET scans. The current diagnostic criteria for amyloid PET include visual assessment of tracer uptake in the gray matter [11]; however, head motion makes it difficult to distinguish between the gray and white matter. Before the respiratory motion correction algorithm was developed for oncology whole-body PET, if the accurate evaluation of lesions in the lung base or upper abdomen was difficult owing to respiratory motion, the relevant area was imaged again. Head motion is unpredictable and prevalent in patients with poor motor control. In addition, the re-scan time would not be shorter than the initial scan time in amyloid brain PET, and the additional radiation exposure from the re-scan must be considered. Therefore, a post-acquisition head motion correction algorithm is necessary.

In this study, we applied an advanced motion correction algorithm to amyloid brain PET images and compared it to conventional methods. When dividing the list-mode file into a certain number of contiguous intervals according to specific criteria, we initially applied traditional K-means clustering to find intervals when the patient was stationary. However, the traditional algorithm was limited by requiring knowledge of the number of clusters in advance and by assigning observations to clusters regardless of their observation time. To overcome these limitations, we developed a modified “Merging Adjacent Clustering” method so that time-continuous intervals could be identified. Therefore, it could calculate the number of clusters without knowing in advance how many there are. In the process of estimation of 3D motion transformations among the various intervals, when the iterations are complete, the final image represents the final floating image, motion corrected to the frame of the reference image. For summing all corrected motion frames into one frame, rather than the typical method of registering all floating images to a single target, we applied a new “Summing Tree Structural Motion Correction” algorithm. This approach could reduce image noise effectively by summing the tree nodes of each image after motion correction. In addition, during the iterative reconstruction process, we calculated a motion correction method, resulting in a final image with spatial resolution nearly identical to that of the reference image acquired without motion.

The visual reading system for amyloid brain PET images used in routine clinical practice is intuitive and easy to use without any special program [11]. In previous studies, interobserver agreement for visual analysis has been reported to be good in well-trained readers (κ = 0.70–0.93) [17, 18]. In this study, there was almost perfect agreement between two nuclear medicine physicians with respect to interobserver assessment in the conventional and motion-corrected PET images (κ = 0.81 and κ = 0.88, respectively). In the regional analysis, higher interobserver agreement was observed in motion-corrected PET images in all specific regions, and there was a statistically significant increase in the left PC/P region (p = 0.038). The PC/P region is one of the earliest brain regions of the brain to be affected by Aβ pathology [19, 20], and it is associated with executive function changes that may precede memory decline in preclinical AD [20]. This region is relatively small compared to the frontal, parietal, and lateral temporal cortices in the transaxial plane and would thus be expected to be more affected by head motion. The increase in interobserver agreement in this area on motion-corrected PET images indicates that this algorithm may help improve the diagnostic performance of amyloid brain PET.

In this study, the final PET interpretations of amyloid deposition changed after motion correction in 11 (10%) of 108 PET image sets. However, the number of equivocal PET results did not change significantly before or after motion correction for either reader (n = 5 ◊ 5 for reader A; n = 4 ◊ 4 for reader B). Amyloid-equivocal results are inevitable in the binary visual assessment of amyloid brain PET images because of anatomical problems, such as severe atrophy, cortical deformity, or encephalomalacia, as well as the low burden of amyloid deposition [21]. However, the amyloid-equivocal results for head motion are expected to improve with the application of the motion correction algorithm. Of the 5 patients (5%) with poor image quality in the conventional PET images, 4 patients showed definite FMM uptake throughout the entire cerebral cortex; therefore, both readers judged them to have amyloid positivity despite severe motion, and these results did not change even after motion correction (Fig. 3). However, the remaining patient was judged to have amyloid equivocality owing to ambiguous uptake in some areas, and the final report of both readers was changed to amyloid negativity after motion correction (Fig. 4).

18F-flutemetamol brain PET study in a 92-year-old woman. Both readers interpreted the conventional PET images (A) as amyloid positive with poor image quality. The motion-corrected PET images (B) had good image quality and were still interpreted as amyloid positive. The graphs showed the rotation (C) and translation (D) estimated by the motion correction algorithm that revealed 39 frequent and large head movements

18F-flutemetamol brain PET images from a 67-year-old man with obvious head movements during scanning. Both readers interpreted the conventional PET images (A) as amyloid equivocal with poor image quality. After motion correction, the preserved gray matter to white matter contrast was clearly revealed on the PET images (B, arrows). Accordingly, both readers changed their interpretation to amyloid negative with good image quality

Quantitative assessment is widely used in 18F-fluorodeoxyglucose (FDG) PET for the staging and evaluation of treatment response in solid tumors. It shows excellent intra- and interobserver reproducibility and can provide more objective and detailed information than visual analysis [22,23,24]. For the evaluation of amyloid brain PET images in patients with cognitive impairment, quantitative assessment is mainly used in research settings [25]. Cholinesterase inhibitors are currently used to treat AD; however, they are not curative and cannot stop disease progression. Recently, monoclonal antibody drugs (e.g., lecanemab or donanemab) that can reduce Aβ accumulation have been actively developed, and accordingly, the role of amyloid and tau PET for quantifying Aβ and tau burden in vivo has been emphasized [26,27,28]. SUVR is a representative value for the quantification of amyloid and tau PET images and is widely used in clinical trials to evaluate the therapeutic effects of new drugs. Although the appropriate reference region and cut-off values for each radiotracer remain controversial, SUVR can be a supplementary parameter for clearer judgments in equivocal PET cases using a visual reading system [8, 29,30,31]. The accurate setting of the VOI along the gray matter is essential for obtaining accurate SUVR from amyloid brain PET. Unlike in oncology PET where the boundary between FDG uptake by the tumor and the background is clear, in amyloid brain PET, the boundary between the gray and white matter can be easily obscured, even by fine motion.

Given that the uptake of white matter is much higher than that of gray matter in patients with amyloid-negative PET findings, improper inclusion of white matter uptake into the VOI due to head motion leads to a false elevation of the SUVR. In this study, 41 (73%) of the 56 conventional PET images with 1 or 2 motion events were visually assessed by both readers as having good image quality. Of the 68 conventional PET images assessed by both readers as having good image quality, 51 images had more than 1 motion event, including 1 image with as many as 17 motion events (Fig. 5). Meanwhile, 1 conventional PET image with 4 motion events was assessed as having visually poor image quality. The reader may recognize that the SUVR in brain PET cases with obvious large head movements is not reliable. However, the values obtained from PET cases with visually undetectable, small head movements may also be unreliable. In this study, the difference in UR between the conventional and motion-corrected PET images was significantly greater in the group with head motion than in the group without head motion. Therefore, head motion during brain PET imaging can affect image interpretation by making quantitative values inaccurate. This problem would be particularly severe in the evaluation of treatment responses by comparing PET images before and after treatment. In the group with head motion, the severity of the SUVR error can vary depending on the degree of amyloid deposition and the magnitude of motion, rather than the number of motions.

18F-flutemetamol brain PET study in a 79-year-old man. Both readers interpreted both conventional (A) and motion-corrected (B) PET images as amyloid positive with good image quality. However, the motion correction algorithm detected 17 subtle head movements with rotation changes of up to 2.39° (C) and translations of up to 3.09 mm (D) relative to the initial position. In quantitative analysis, the UR differed by 0.12 between the conventional and motion-corrected PET images. The difference in normalized SUV between both PET images was identified in the difference map by subtracting the conventional from the motion-corrected PET images (E)