Abstract
Structural magnetic resonance (MR) images concomitantly acquired with PET images can provide crucial anatomical information for precise quantitative analysis. However, in the clinical setting, not all the subjects have corresponding MR. Here, we developed a model to generate structural MR images from amyloid PET using deep generative networks. We applied our model to quantification of cortical amyloid load without structural MR. Methods: We used florbetapir PET and structural MR data of Alzheimer’s Disease Neuroimaging Initiative database. The generative network was trained to generate realistic structural MR images from florbetapir PET images. After the training, the model was applied to the quantification of cortical amyloid load. PET images were spatially normalized to the template space using the generated MR and then standardized uptake value ratio (SUVR) of the target regions was measured by predefined regions-of-interests. A real MR-based quantification was used as the gold standard to measure the accuracy of our approach. Other MR-less methods, a normal PET template-based, multi-atlas PET template-based and PET segmentation-based normalization/quantification methods, were also tested. We compared performance of quantification methods using generated MR with that of MR-based and MR-less quantification methods. Results: Generated MR images from florbetapir PET showed visually similar signal patterns to the real MR. The structural similarity index between real and generated MR was 0.91 ± 0.04. Mean absolute error of SUVR of cortical composite regions estimated by the generated MR-based method was 0.04±0.03, which was significantly smaller than other MR-less methods (0.29±0.12 for the normal PET-template, 0.12±0.07 for multiatlas PET-template and 0.08±0.06 for PET segmentation-based methods). Bland-Altman plots revealed that the generated MR-based SUVR quantification was the closest to the SUVR values estimated by the real MR-based method. Conclusion: Structural MR images were successfully generated from amyloid PET images using deep generative networks. Generated MR images could be used as template for accurate and precise amyloid quantification. This generative method might be used to generate multimodal images of various organs for further quantitative analyses.
- Neurology
- PET
- Deep learning
- Florbetapir PET
- Generative adversarial network
- MR generation
- PET quantification
- Copyright © 2017 by the Society of Nuclear Medicine and Molecular Imaging, Inc.