V3I2P31

Integrating MRI, Clinical, and Genetic Features Using a Dual-Attention CNN-MLP for Alzheimer’s Diagnosis

Dr. V. Maniraj^1*, K. Vaishali²

Abstract

Alzheimer’s disease (AD) is a progressive and irreversible neurodegenerative disorder, and mild cognitive impairment (MCI) is widely recognized as an intermediate stage preceding AD. Early and accurate identification of AD is crucial for timely prevention and intervention. Although numerous computational approaches have been developed for AD diagnosis, many rely solely on neuroimaging data and overlook additional informative modalities such as genetic and clinical data. Moreover, the interpretability of some existing models remains limited.

To address these issues, we propose a novel dual-attention CNN–MLP framework, termed DANMLP, designed for computer-aided AD diagnosis through the integration of structural MRI (sMRI), clinical variables (including demographic and neuropsychological measures), and APOE genetic information. The DANMLP model is composed of four main components:

(1) a Patch-CNN module for extracting local image features from sMRI patches;

(2) a position self-attention mechanism for modeling intra-patch feature dependencies;

(3) a channel self-attention mechanism for capturing inter-patch feature relationships; and

(4) two multilayer perceptrons for clinical feature extraction and final diagnostic classification.

Experimental results on the ADNI dataset using 5-fold cross-validation demonstrate that DANMLP achieves classification accuracies of 93% for AD vs. MCI and 82.4% for MCI vs. NC, outperforming five state-of-the-art methods by 0.2%–15.2% and 3.4%–26.8%, respectively. Furthermore, individualized visualization of salient brain regions enhances the model’s interpretability and provides valuable assistance for early clinical assessment. Overall, these findings indicate that DANMLP is an effective and reliable tool for diagnosing AD and MCI.

Keywords:

Object segmentation, Diagnostic imaging, Fuzzy inference, Graph-based image segmentation, Hierarchical neural learning, U-shaped segmentation network, MR imaging, Biological data analysis.