An Adaptive Feature Fusion Network for Alzheimer's Disease Prediction.

Structural Magnetic Resonance Imaging (sMRI) of brain structures has proven effective in predicting early lesions associated with Alzheimer’s disease (AD). However, identifying the AD lesion area solely through sMRI is challenging, as only a few abnormal texture areas are directly linked to the lesion. Moreover, the observed lesion area varies when examining two-dimensional MRI slides in different directions within three-dimensional space. Traditional convolutional neural networks struggle to accurately focus on the AD lesion structure. To address this issue, we propose an adaptive feature fusion network for AD prediction. Firstly, an adaptive feature fusion module is constructed to enhance attention towards lesion areas by considering features from three directions and fusing them together. Secondly, a multi-channel group convolution module is designed to improve the network’s ability to extract fine-grained features by separating convolutional channels. Finally, a regularized loss function is introduced to combine the SoftMax loss function and clustering loss function. This helps enhance the network’s ability to differentiate between different sample types. Experimental results from binary classification tests on the dataset obtained from Alzheimer’s Disease Neuroimaging Initiative (ADNI) demonstrate that our proposed method accurately distinguishes between normal control (NC), progressive mild cognitive impairment (pMCI), stable mild cognitive impairment (sMCI), and AD.