Alzheimer’s disease (AD) is a complex condition that currently lacks definitive imaging diagnostic gold standards. In clinical practice, physicians typically integrate cognitive scores with imaging examination results for comprehensive diagnosis. To fully exploit the potential of structural Magnetic Resonance Imaging (sMRI) and enhance its application value in AD diagnosis, this study proposes a novel methodological framework. First, a cognitive score prediction model and an image feature learning model based on sMRI are constructed. Subsequently, the predicted cognitive scores are integrated with the image feature through a designed multimodal fusion strategy. Our built end-to-end framework utilizes regression tasks based on real cognitive scores during the training phase to assist in optimizing the representation of image features, thereby allowing AD diagnosis solely based on imaging data during the testing phase. By integrating cognitive phenotypic information derived from imaging data, the proposed method addresses the limitations of single-modality imaging and improves diagnostic performance on the ADNI dataset. Meanwhile, it avoids relying on clinical phenotype labels during the diagnostic decision process, thereby preserving the physical interpretability of the image-based diagnosis method.