Cascaded deep learning enables multimodal brain PET spatial normalization and quantification for Alzheimer's disease.
Cheng Tang1, Anqi Tang2, Mengyu Wan2
1Department of Nuclear Medicine, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, 430022, China; First School of Clinical Medicine, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, 430022, China.
Neuroimage
|June 6, 2026
Summary
A new Deep Cascaded Cerebral Calculator (DCCC) enables accurate Alzheimer's disease (AD) biomarker analysis using only PET scans, removing the need for MRI. This automated tool speeds up diagnosis and research, making AD imaging more accessible.
Area of Science:
- Neuroimaging
- Medical Imaging Analysis
- Artificial Intelligence in Medicine
Background:
- Semi-quantitative positron emission tomography (PET) analysis is crucial for Alzheimer's disease (AD) diagnosis and research.
- Current quantification methods often require structural MRI for spatial normalization or use computationally intensive software, limiting clinical application.
- There is a need for efficient, MRI-independent PET quantification methods.
Purpose of the Study:
- To develop and validate the Deep Cascaded Cerebral Calculator (DCCC), a fully automated, PET-only framework for Alzheimer's disease neuroimaging analysis.
- To enable rapid spatial normalization and accurate biomarker quantification without the need for structural MRI.
- To facilitate large-scale, tracer-agnostic analyses in AD research.
Main Methods:
- Development of the Deep Cascaded Cerebral Calculator (DCCC), a PET-only framework utilizing cascaded CNN-based rigid/affine and VoxelMorph-based elastic registration for spatial normalization.
- Retrospective, multi-center study involving 3,539 patients and 6,531 scans across 7 modalities and 13 tracers.
- Benchmarking DCCC against the standard MRI-guided SPM12 pipeline and other PET-only tools using meta-region-of-interest (ROI) Standard Uptake Value ratio (SUVr) and correlation analyses.
Main Results:
- DCCC achieved high accuracy in SUVr quantification (mean absolute relative error of 1.34±0.59%) and strong correlation (Pearson correlation of 0.96±0.02) with standard methods.
- The framework demonstrated robust generalization to unseen tracers and modalities, including neuroinflammation and methionine metabolism imaging.
- Centiloid and CenTauRz estimates derived from DCCC were highly accurate (R²>0.97) with rapid processing speeds (1.22±0.64 s per image).
- DCCC showed utility in longitudinal tracking, deep learning preprocessing for classification, and clinical support, with metrics adopted in a significant percentage of Aβ and tau cases.
Conclusions:
- The Deep Cascaded Cerebral Calculator (DCCC) provides accurate, PET-only standardization for Alzheimer's disease biomarker estimation, eliminating the need for MRI.
- DCCC facilitates harmonized biomarker quantification and enables large-scale, tracer-agnostic analyses, enhancing clinical accessibility and research capabilities.
- The availability of a free standalone program and 3D Slicer plugin promotes widespread adoption and application in AD neuroimaging.
