The auxiliary diagnostic methods for Alzheimer’s disease (AD) using single-modal data suffer from a lack of professionally annotated imaging data, unstable feature extraction and high requirements for computing power. A multimodal lightweight Alzheimer’s disease auxiliary diagnostic network (LightMoDAD) was proposed, and magnetic resonance imaging (MRI), positron emission tomography (PET) imaging data, and psychometric score data were utilized in the network. In the image feature extraction module, spatial and channel reconstruction convolution was employed to extract local features, global filtering to extract global features, and feature registration and addition to achieve multimodal feature fusion. In the text feature extraction module, the features extracted by separable convolution from psychometric score data were integrated with multimodal image feature, and transfer learning was used to improve feature discrimination. A multi-layer perceptron was applied to recognize complex patterns and features, and the classification accuracy of the proposed network was improved. Experiments were conducted on the ADNI dataset, and LightMoDAD’s accuracy, sensitivity and specificity were 0.980, 0.985 and 0.975, respectively. Experimental results show that the proposed network enhances physicians’ diagnosis efficiency and holds potential for mobile platform deployment.