Abstract
1779
Objectives: We aimed to make use of non-local mean (NLM) and median (NLEM) methods, with and without anatomy information, to significantly enhance quality and quantitative accuracy of PET images, with special focus on amyloid imaging.
Methods: Using anatomy information to enhance PET images has been a popular strategy, but many previous works have incorporated too much anatomy information, biasing PET images. In this study, we only use the anatomy regions as masks to restrict the search window in non-local type methods. We investigated two methods, namely non-local means (NLM) and non-local Euclidean medians (NLEM): NLM and NLM-A compute the mean of the patches in the search window to update the center pixel of the patch in the original image, while NLEM and NLEM-A use the Euclidean median. To further improve the PET images, we also implemented the anatomy-incorporated counterparts (NLM-A and NLEM-A), utilizing the anatomical region masks to restrict the search window, thus restricting usage of non-local information to voxels that belong to the same tissue. We set the overall search window size to 21[asterisk]21 and patch size to 7[asterisk]7 within a given slice. We performed two studies to evaluate these four methods. First, we utilized the classical NEMA NU-2 image quality phantom with six spheres of different diameters (10, 13, 17, 22, 28 and 37 mm). Second, we subsequently performed realistic amyloid PET simulations based on real data from human amyloid dynamic PET (11C-PIB) and MRI studies. We labeled regions including GM, WM, CSF, caudate, putamen and thalamus, as areas of interest for Alzheimer's disease research. We incorporated realistic clinical data activity distributions in our simulation studies based on participant images.
Results: Compared with the no-anatomy versions, anatomy-guided methods performed better in both studies, specifically in visual image and image roughness (spatial variance) performance. At matched noise levels, anatomy-based non-local methods (both NLM-A and NLEM-A) significantly lowered bias. For example, in GM of NLEM-A, at matched 20% noise (image roughness), the bias was reduced by 88%, and at 25% noise, the bias was lowered by 66%, and at 35% noise, the bias was lowered by 48%. The ensemble variance of these four methods looked relatively matched. In general, the effectiveness and trend of the four methods are not greatly affected by noise levels. For high noise levels, NLEM and NLEM-A perform somewhat more effectively than NLM and NLM-A in some regions (e.g. thalamus, WM), because the median operator (instead of mean operator) is less impacted by the presence of outliers. Furthermore, without anatomy guidance, NLM and NLEM use information from a very wide search window to update the image, and this can result in additional bias in the resulting image estimates. Conclusion: Overall, we have demonstrated that non-local methods in combination with anatomic guidance have a significant potential to improve quality and quantitative accuracy of amyloid PET images.
In this issue
