Fast and accurate amyloid brain PET quantification without MRI using deep neural networks

Abstract

AbstractBackgroundWe propose a novel method for automatic quantification of amyloid PET using a deep learning‐based spatial normalization (SN) of PET images, which does not require corresponding MRI or CT image of the same patient. In this study, the accuracy of the proposed method was evaluated for three different amyloid PET radiotracers by comparison with MRI parcellation‐based PET quantification using FreeSurfer, which is more accurate than SN‐based method but requires significantly longer computation time.MethodDeep neural network model used for SN of amyloid PET images was trained using 994 multicenter amyloid PET images, as well as corresponding 3D MRIs of the patients with Alzheimer’s disease and mild cognitive impairments and cognitive normal subjects. The accuracy of SN and SUVR quantification accuracy relative to FreeSurfer‐based estimation was evaluated using other 148 PET images. Additional external validation was also performed using independent data set (30 18F‐Flutemetamol, 67 18F‐Florbetaben, and 39 18F‐Florbetapir). For the comparison, PET SN was also conducted using SPM12 program. Then, the quantification results using SN methods were compared with SUVR values obtained using on FreeSurfer in individual brain space. Reference region was cerebellar grey matter and AAL3 atlas was used to extract the regional SUVR values.ResultsThe quantification results using the proposed method showed stronger correlations with FreeSurfer estimates than SPM SN using MRI did. For example, the slope, y‐intercept and R2 value between SPM and FreeSurfer for global cortex were 0.869, 0.113 and 0.946, respectively. On the other hand, those for proposed method were 1.019, ‐0.016 and 0.986. The external validation study also demonstrated the better performance of proposed method without MR images than SPM with MRI. In most brain regions, it outperformed the SPM SN in terms of the linear regression parameters and intraclass correlation coefficients.ConclusionIn this study, we evaluated a novel deep learning‐based SN method, which allows quantitative analysis of amyloid brain PET images without structural MRI. The proposed quantification results showed strong correlation with MRI‐parcellation‐based amyloid PET quantification using FreeSurfer in all clinical amyloid radiotracers. Therefore, the proposed method will be useful for investigating Alzheimer’s disease and related brain disorders using amyloid PET scans.