A multi-view DTI feature fusion framework for enhanced diagnosis of Alzheimer's disease.

Alzheimer's disease (AD) is an irreversible neurodegenerative disorder. Diffusion tensor imaging (DTI) is widely used to detect brain alterations for diagnosis, but most methods rely on single-scale information. Therefore, this study proposes the multi-view feature learning framework incorporating residual block-based 3D convolutional neural network (3D-CNN) for AD diagnosis. First, tract-based spatial statistics were applied to extract voxel-based features from fractional anisotropy (FA) and mean diffusivity (MD) maps. Second, the residual block-based 3D-CNN model was employed to extract high-level deep features, enhancing model ability to capture global contextual information. Third, fiber tracking was used to construct structural connectivity networks, which served as connectivity-based features. Fourth, radiomics was applied to extract texture and shape features from FA and MD images. These four types of features were linearly combined and subsequently reduced in dimensionality using the ReliefF algorithm. Finally, an ensemble learning strategy was employed to perform three binary classification tasks among the AD, mild cognitive impairment (MCI), and normal control (NC) groups. Additionally, layer-wise relevance propagation (LRP) was utilized to improve the interpretability of the 3D-CNN model. Evaluated on 427 subjects from the Alzheimer's Disease Neuroimaging Initiative database, the framework integrates complementary multi-scale information, achieving superior performance. For the AD vs. NC classification, it attained an accuracy of 97.6%, a sensitivity of 98.0%, and an area under the curve of 0.964, outperforming several state-of-the-art methods. These results demonstrate that our approach enhances diagnostic accuracy and contributes to understanding disease mechanisms by identifying multi-scale biomarkers associated with known AD pathology.