Brain fluorodeoxyglucose PET anatomical segmentation via AI: extensive validation in the neurodegenerative spectrum.

The semi-quantitative assessment of brain [18F]FDG PET provides a more objective interpretation and improved accuracy in differentiating across neurodegenerative diseases. However, the correct identification of anatomical regions of interest without a structural MRI is challenging. Thus, this study aims to develop a deep-learning-based (DL-based) model for the automatic segmentation of 52 anatomical regions in brain [18F]FDG PET images and validate it across different metabolic profiles. 1628 brain [18F]FDG PET images of 1099 subjects were included in the internal dataset, comprising cognitively normal subjects (n=537), and patients with mild cognitive impairment (n=538), subjective memory concerns (n=70), Alzheimer's disease (n=330), frontotemporal lobar degeneration (n=95) and Lewy body dementia (n=58). Train-test split yielded 1109/519 images (631/468 subjects) for training and internal testing of a DL-based model, respectively. An additional dataset of 108 [18F]FDG PET images was included for external validation. Ground-truth segmentation was performed on the paired T1-weighted MRI image for each [18F]FDG PET image, for a total of 52 anatomical regions of interest. Atlas-based segmentation was used as a benchmark. The Dice similarity coefficient (DSC) was used to assess segmentation performance. Agreement in mean pons-based standardised uptake value ratio (SUVRmean) quantification was assessed through the intraclass correlation coefficient (ICC) and relative deviation in absolute value. Per-region mean DSC for the DL-based segmentations ranged from 0.729 to 0.923 in the internal test set. Global mean DSC was 0.84±0.06. SUVRmean quantification of the different regions using the DL-based segmentation masks showed strong agreement with that obtained using the ground-truth segmentation masks, with an average ICC of 0.96±0.02. The per-region average of relative SUVRmean deviation did not exceed 5%. DL-based segmentation significantly outperformed atlas-based segmentation (p<0.05). Similar segmentation performance was obtained in the external validation dataset. DL-based anatomical segmentation of brain [18F]FDG PET proved to be robust across a wide spectrum of neurodegenerative diseases. Semi-quantitative assessment was comparable with that obtained with MRI-based segmentation. DL-based segmentation, therefore, proved to be a reliable alternative when MRI isn't available, and is a better option than the commonly used atlas-based approach.