Abstract
In recent studies on Alzheimer's disease (AD), various network models have shown significant potential in disease prediction. However, traditional CNNs often rely on parameterized loss functions, limiting the robustness of these models. Additionally, their high computational complexity increases resource demands. To address these challenges, this study proposes a novel prediction model that integrates the strengths of ViTs and the MedMamba module. First, this study draws on the SS-Conv-SSM module from the MedMamba model, which processes image branches in parallel to extract richer and more refined features. Building on this, we optimized the original purely convolutional structure into a hybrid architecture combining convolution and Transformer layers. This not only reduces the computational burden and enhances operational efficiency but also improves the model's ability to capture global features. Moreover, we introduced a new self-attention mechanism into the model's MDTA module, reducing the computational complexity from quadratic to linear. This allows the model to maintain high performance while achieving more lightweight and efficient operations. The final experimental results demonstrate that this model outperforms current state-of-the-art methods in predicting Alzheimer's using brain [Formula: see text]-FDG PET (fluorodeoxyglucose positron emission tomography) images, particularly excelling in distinguishing AD from mild cognitive impairment.