Cascaded Deep Learning enables multimodal brain PET spatial normalization and quantification for Alzheimer's disease.
Authors
Affiliations (5)
Affiliations (5)
- Department 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.
- Department of Nuclear Medicine, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, 430022, China; Hubei Key Laboratory of Molecular Imaging, Wuhan, 430022, China; Key Laboratory of Biological Targeted Therapy, the Ministry of Education, Wuhan, 430022, Hubei Province, China.
- Tsinghua University, Beijing, 100084, China.
- Department of Nuclear Medicine, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, 430022, China; Hubei Key Laboratory of Molecular Imaging, Wuhan, 430022, China; Key Laboratory of Biological Targeted Therapy, the Ministry of Education, Wuhan, 430022, Hubei Province, China. Electronic address: [email protected].
- Department of Nuclear Medicine, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, 430022, China; Hubei Key Laboratory of Molecular Imaging, Wuhan, 430022, China; Key Laboratory of Biological Targeted Therapy, the Ministry of Education, Wuhan, 430022, Hubei Province, China. Electronic address: [email protected].
Abstract
Semi-quantitative positron emission tomography (PET) analysis, particularly Centiloid and CenTauRz scaling, is essential for Alzheimer's disease (AD) research and diagnosis. However, standard quantification workflows often depend on structural MRI for spatial normalization (SN) or rely on computationally intensive software, limiting clinical accessibility. In this retrospective, multi-center study (3,539 patients; 6,531 scans; 2005-2025), we compiled data across 7 modalities and 13 tracers to develop and validate the Deep Cascaded Cerebral Calculator (DCCC). This fully automated, PET-only framework employs cascaded CNN-based rigid/affine and VoxelMorph-based elastic registration modules for rapid SN. We benchmarked 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. DCCC achieved a mean absolute relative SUVr error of 1.34±0.59% and a Pearson correlation of 0.96±0.02, demonstrating robust generalization to unseen tracers and modalities including neuroinflammation and methionine metabolism imaging, with superior consistency compared to conventional template-based PET-only methods. Centiloid and CenTauRz estimates were highly accurate (R²>0.97) with a processing speed of 1.22±0.64 s per image. We further demonstrated DCCC's utility across 3 scenarios: (1) longitudinal tracking, where it identified a distinct low-Centiloid AD subgroup; (2) deep learning preprocessing, yielding classification AUCs comparable to standard methods (P=0.36); and (3) exploratory clinical support, where DCCC-derived metrics were adopted in 79% Aβ and 61% tau cases and were associated with changes in interpretation and increased agreement with reference labels in a multi-reader survey. Collectively, DCCC provides accurate, PET-only standardization, facilitating harmonized biomarker estimation without MRI and enabling large-scale, tracer-agnostic analyses in AD neuroimaging. A free standalone command-line interface program and a 3D Slicer plugin are provided.