Alzheimer’s disease (AD) is a prevalent neurodegenerative disorder that severely impairs the cognitive function and quality of life of patients. Its incidence is escalating with the intensification of the aging society. In this paper, a novel self-supervised contrastive learning framework (CLFAD) is put forward to address the issue of the substantial demand for large-scale labeled data in image classification. We initially employed unlabeled data for contrastive learning pretraining, integrating data augmentation techniques and an adaptive temperature control approach to enhance feature learning, thereby significantly enhancing the model’s performance on unlabeled datasets. Subsequently, we evaluated the performance of the model in diverse AD classification tasks. The experimental outcomes demonstrate that CLFAD outperforms the current mainstream contrastive learning methods in various key indicators, such as accuracy, precision, recall. Additionally, to guarantee the reliability of the results, we implemented a rigorous segmentation strategy and successfully precluded data leakage. This study reveals that applying self-supervised contrastive learning to AD image classification is an effective and feasible tactic, which offers significant support for the early detection and diagnosis of AD.