aceBrain fluorodeoxyglucose PET anatomical segmentation via AI: extensive validation in the neurodegenerative spectrum
Abstract: 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…
TORNAI-SpectrumOfMedicalImaging by Martin Tornai. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0
By September 2026, a peer-reviewed study in Brain reported development and validation of a deep-learning model that automatically segments 52 anatomical regions on brain [18F]FDG PET images without requiring a paired MRI, tested on 1628 images spanning cognitively normal subjects and patients with mild cognitive impairment, Alzheimer's disease, frontotemporal lobar degeneration and Lewy body dementia.
The finding matters because semi-quantitative FDG PET assessment normally depends on MRI for anatomical definition, limiting use when MRI is unavailable; the model offers a reliable alternative that outperformed atlas-based methods, though real-world clinical impact, workflow integration, and performance in unrepresented populations remain to be established.
- Internal dataset included 1628 brain [18F]FDG PET images from 1099 subjects: cognitively normal, mild cognitive impairment, subjective memory concerns, Alzheimer's disease, frontotemporal lobar degeneration and Lewy body dementia.
- Train-test split was 1109/519 images for training and internal testing, plus 108 images for external validation; ground-truth segmentation was performed on paired T1-weighted MRI.
- Per-region mean DSC ranged from 0.729 to 0.923 in internal test set with global mean DSC 0.84±0.06 and average ICC of 0.96±0.02 for SUVRmean agreement.
Deep-learning model automatically segmented 52 anatomical regions on brain FDG PET alone and achieved MRI-comparable pons-based SUVR quantification across cognitively normal and neurodegenerative patients.
The rundown
Researchers trained and tested a deep-learning model to segment 52 anatomical regions directly on brain [18F]FDG PET without structural MRI, using 1628 images from 1099 subjects and an additional 108 images for external validation, with ground truth derived from paired T1-weighted MRI and atlas-based segmentation as benchmark.
In internal testing, per-region mean Dice similarity coefficient ranged from 0.729 to 0.923 with global mean 0.84±0.06, SUVRmean quantification showed average ICC 0.96±0.02 and per-region relative deviation under 5%, and performance was significantly better than atlas-based segmentation with similar results in external validation.
Sources
- Peer-reviewedBrain2026-09-15
How should this claim be treated?
The debate