Alzheimer’s disease (AD) is characterized by pronounced spatial heterogeneity and complex neurodegenerative patterns, which pose significant challenges for representation learning on three-dimensional brain images. Conventional convolutional neural networks relying on regular grid sampling struggle to align localized structural degeneration, and their performance is further compromised by class imbalance and a high proportion of weakly discriminative samples, leading to suboptimal optimization dynamics and reduced generalization ability. To address the aforementioned challenges, this study proposes Deformable Attention and Risk-aware 3D Network (DAR-3DNet), a modeling framework for 3D magnetic resonance imaging (MRI) classification. The proposed method incorporates deformable sampling and sampling-point modulation into spatial attention generation, enabling the attention estimation process to better adapt to the non-rigid spatial patterns associated with brain structural degeneration. On this basis, an instance-adaptive label smoothing loss with composite risk, termed Instance-wise Adaptive Label Smoothing Loss with Composite Risk (IASLCR), is further introduced to dynamically adjust supervision strength based on sample-specific risk, thereby alleviating optimization bias caused by class imbalance and weakly discriminative samples. Experiments conducted on 1749 structural MRI scans from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset show that the proposed method achieves accuracy values above 0.93 on the AD vs. normal control (NC), mild cognitive impairment (MCI) vs. NC, and AD vs. MCI classification tasks, while yielding better overall performance than the evaluated baseline models. These results suggest that the proposed framework has considerable potential for structural MRI-based AD/MCI classification.
Bi et al. (Fri,) studied this question.