This study proposes a multimodal diagnosis model for Alzheimer's disease (AD) based on improved Transformer and 3D convolutional neural network (3DCNN). The model uses structural magnetic resonance imaging (sMRI) and positron emission tomography (PET) data to extract deep features through 3DCNN. An improved Transformer is then used to learn global correlation information among features, followed by fusion of information from different modalities for identification. A model-based visualization method is used to explain the model's decisions and identify brain regions related to AD. The model achieved a classification accuracy of 98.1% on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. Visualization results showed distinct brain regions associated with AD diagnosis across different image modalities, with the left parahippocampal region consistently emerging as a significant area. The model underwent extensive experiments, demonstrating high performance and reliability. The visualization method based on the model's characteristics improved interpretability and identified disease-related brain regions, providing reliable information for AD research. The model combines the strengths of 3DCNN and Transformer to extract and fuse features from multimodal medical images. It uses an improved Transformer to learn global correlation information, which enhances the model's performance. The model's results show that multimodal data outperforms single-modal data in AD diagnosis. The model's visualization analysis helps identify key brain regions related to AD, contributing to a better understanding of the disease's pathogenesis. The study highlights the importance of multimodal data in AD diagnosis and the effectiveness of deep learning techniques in extracting meaningful features. The model's results suggest that AD-related brain disorders can be precisely examined using multimodal medical images and deep learning. The model's visualization method provides additional insights for clinical diagnosis and research. The study contributes to the development of more accurate and interpretable AD diagnosis models.This study proposes a multimodal diagnosis model for Alzheimer's disease (AD) based on improved Transformer and 3D convolutional neural network (3DCNN). The model uses structural magnetic resonance imaging (sMRI) and positron emission tomography (PET) data to extract deep features through 3DCNN. An improved Transformer is then used to learn global correlation information among features, followed by fusion of information from different modalities for identification. A model-based visualization method is used to explain the model's decisions and identify brain regions related to AD. The model achieved a classification accuracy of 98.1% on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. Visualization results showed distinct brain regions associated with AD diagnosis across different image modalities, with the left parahippocampal region consistently emerging as a significant area. The model underwent extensive experiments, demonstrating high performance and reliability. The visualization method based on the model's characteristics improved interpretability and identified disease-related brain regions, providing reliable information for AD research. The model combines the strengths of 3DCNN and Transformer to extract and fuse features from multimodal medical images. It uses an improved Transformer to learn global correlation information, which enhances the model's performance. The model's results show that multimodal data outperforms single-modal data in AD diagnosis. The model's visualization analysis helps identify key brain regions related to AD, contributing to a better understanding of the disease's pathogenesis. The study highlights the importance of multimodal data in AD diagnosis and the effectiveness of deep learning techniques in extracting meaningful features. The model's results suggest that AD-related brain disorders can be precisely examined using multimodal medical images and deep learning. The model's visualization method provides additional insights for clinical diagnosis and research. The study contributes to the development of more accurate and interpretable AD diagnosis models.