Deep Learning–Based Multimodal Fusion of Wavelet-Decomposed MRI and PET for Alzheimer’s Disease Diagnosis

Early and accurate diagnosis of Alzheimer’s disease (AD) is crucial for timely clinical intervention and treatment planning. Recent advances in deep learning have facilitated the automatic extraction of informative biomarkers from neuroimaging data. In this study, we proposed a multimodal framework that integrates structural magnetic resonance imaging (MRI) and functional positron emission tomography (PET) using three-dimensional convolutional neural networks (3D-CNNs) with wavelet-domain feature extraction (CO.2). First, MRI and PET volumes are decomposed into multiple sub-bands via discrete wavelet transform (DWT) to capture complementary frequency-domain patterns. Next, a custom 3D-CNN extracts multi-scale hierarchical features from each modality (CO.3,4). The resulting modality-specific embeddings are fused at the decision level through a stacked one-dimensional CNN (1D-CNN) ensemble to classify subjects into AD, mild cognitive impairment (MCI), and normal control (NC) groups. Experiments on the ADNI dataset show strong performance across three binary tasks, achieving 97.9% (AD vs. NC), 97.5% (MCI vs. NC), and 99.6% (AD vs. MCI, AUC ≈ 0.997). These results demonstrate that the proposed wavelet-enhanced stacked multimodal model provides a robust and generalizable tool for early-stage AD detection, particularly in distinguishing subtle transitions between MCI and AD (CO.2).