Fast and accurate spatial normalization of amyloid PET images without MRI using deep neural networks

1403 Objectives: Accurate spatial normalization (SN) of amyloid PET images for Alzheimer’s disease assessment without co-registered anatomical magnetic resonance imaging (MRI) of the same individual is technically challenging due to the different uptake patterns depending on the degree of disease progression. Recent studies have shown that fast image registration methods based on deep neural networks have great potential to replace conventional analytic algorithms. Here, we propose a deep neural network-based spatial normalization method for amyloid PET images that do not require corresponding individual MRI. Methods: We trained and tested the deep convolutional neural network model consisting of three cascaded U-Nets using 706 pairs of simultaneously acquired 11C-PIB PET and T1-weighted 3D MRI scans of AD, MCI, and cognitively normal subjects (training=553, test=153). The proposed network takes an affine-registered amyloid PET image as an input and generates a local displacement field for nonlinear registration. The generated displacement field is applied to co-registered MRI in the training phase, and cross-correlation loss between the spatially normalized MRI and T1 template (individual MNI152) was minimized with error backpropagation. Additionally, the grey matter (GM) the segment of each MRI was provided to the network as input and deformed by the generated displacement field. Dice loss was calculated between the deformed GM segment and GM of the MNI152 template, which was minimized together with cross-correlation loss. Spatially normalized PET images were not required as a reference in the training phase, and only PET images were used to create the deformation field and normalized images in the test phase. SN results were compared with the SPM12 SN output which required a co-registered MRI. FreeSurfer was used to segment the detailed GMs in individual subject spaces. SUVR values for predefined ROIs were calculated using cerebellum grey as a reference region. The SUVR values obtained from individual space were ground truth and compared to the SUVRs obtained from SN images using SPM12 and the proposed method. Results: After training, the proposed network successfully generated displacement fields and yielded accurate spatially normalized PET images. SPM12 SN was not accurate enough for patients with severe brain atrophy. However, the proposed network performed well for these patients. The average values of mutual information for the test data were 1.18, 1.21, and 1.67 for SPM12 SN, the proposed network without GM segments, and the proposed network with GM segments, respectively. In addition, the proposed network with GM segments showed better linearity than the SPM12 SN in various GM ROIs. The superiority of the proposed method over the SPM12 was most pronounced in the deep grey ROIs. The computation times were approximately 30 seconds and 1 second for SPM12 and the proposed network, respectively. Conclusion: The proposed method has great potential for routine analysis of amyloid PET images in clinical practice and research because it does not require 3D MRI for the SN of PET images and has a short computation time.