Hierarchical contrastive learning for multi-modal representation in Alzheimer’s disease diagnosis

Alzheimer’s disease (AD) diagnosis increasingly relies on the integration of multi-modal data, including structural magnetic resonance imaging, positron emission tomography and cerebrospinal fluid biomarkers. However, existing methods often inadequately explore the discriminative information within each modality and fail to fully capture complementary interactions between modalities. In this paper, we propose a hierarchical contrastive learning method which can fully utilize the intra-modal and inter-modal information, guiding the model to learn a well-structured feature space to improve the performance of diagnosis. Specifically, our method first introduces intra-modal alignment, which enforces supervised contrastive constraints within each modality, thereby effectively exploiting discriminative information between different clinically diagnosed groups. Secondly, we innovatively propose a clinical label-guided inter-modal alignment method to exploit the complementary interactions between modalities, leading to more discriminative and coherent representations. Furthermore, a dynamic learning strategy is employed between the intra-modal and inter-modal alignments to ensure the stability of the training process and achieve the comprehensive hierarchical contrastive learning. Experimental results on multiple datasets demonstrate that our method outperforms current state-of-the-art approaches, validating its efficiency in learning discriminative multi-modal representations for AD related diagnosis. The source code is available at https://github.com/qizhiJing/HSCL .