Cross-Domain Adaptive Fusion Network (CDAFN) with RAD2F, WEDFE, and TGAF Feature Fusion for Automated Alzheimer's Disease Classification from Brain MRI

Abstract

Despite advances in magnetic resonance imaging (MRI) technologies and the arrival of deep learning approaches, early and accurate diagnosis of Alzheimer's disease (AD) is still a difficult task, as subtle changes in the neuroanatomical structure of the brain, inter-subject variability, and the lack of the ability to interpret only deep learning models make the process challenging. In this paper, we present a novel multi-stage hybrid framework that combines anatomical, spectral, and texture information to enhance the interpretability of the clinical scene while simultaneously boosting diagnostic accuracy. The proposed framework is comprised of four key elements. First, the Region-Aware Dual-Domain Feature Fusion (RAD²F) module fuses handcrafted Gray-Level Co-occurrence Matrix (GLCM) features with deep features from ResNet-50 on anatomically important brain regions. Second, the Wavelet-Enhanced Deep Feature Embedding (WEDFE) module generates complementary multi-scale spectral and spatial features through Discrete Wavelet Transform further enhancing the MRI acquisition-related and noise robustness. Third, the Texture-Guided Attention Fusion (TGAF) module introduces texture cues such as GLCM and Local Binary Pattern (LBP) to direct the network towards subtle structural abnormalities linked to Alzheimer's disease. Lastly, the Cross-Domain Adaptive Fusion Network (CDAFN) and the Weighted Multi-Stage Adaptive Selection (WMAS) strategy is applied to filter out features that are redundant, unstable and less discriminative using a genetic algorithm before the classification by a Bayesian-Optimized Random Forest (BO-RF) classifier. A four-class algorithm for Alzheimer's Disease classification was used to evaluate the proposed framework on the OASIS-3 data set. The held-out test set yielded 97.2% accuracy and five-fold scan-level cross-validation yielded 95.72% ± 0.38% accuracy. The strict subject-wise cross-validation resulted in 94.21% ± 0.60% accuracy and 93.88% ± 0.60% weighted F1 score. Some previous work claims greater top-level accuracy but the published work uses different subsets of the OASIS, binary classification problems or other evaluation protocols, so that direct comparison with previous studies is limited because they use different OASIS subsets, class definitions, subject partitions, and evaluation protocols. The proposed framework is competitive under the specific OASIS-3 setting, and shows high interpretability by combining anatomical, spectral and texture features. The framework shows great potential to be a computer-based clinical decision support tool, which needs to be tested clinically in the future and validated externally.