Learning global and local representations for generalized category discovery in neurodegenerative diseases

Abstract Background Artificial intelligence (AI) has achieved notable success in closed‐set classification of neurodegenerative diseases, but its performance degrades significantly in more realistic open‐set settings, such as discovering rare or novel disease subtypes from unlabeled data that are not seen during training. Purpose This study introduces, to our knowledge, the first framework for generalized category discovery (GCD) in neurodegenerative diseases using MRI. It aims to leverage labeled data from known categories to automatically identify novel categories within unlabeled datasets. Methods We propose a GCD framework integrating semi‐supervised contrastive learning with global‐local representation learning to jointly optimize feature learning. Globally, we enforce consistency between weakly augmented views of the same instance by aligning their similarity relationships with other embeddings in a support queue. Locally, we treat neighborhood embeddings from the support queue as positive pairs for cross‐instance contrastive learning, enhancing fine‐grained discrimination. This dual strategy results in a unified learning objective that is applicable to both labeled and unlabeled data, effectively narrowing the representation gap among semantically similar instances. Results We evaluate our method on a benchmark dataset comprising six neurodegenerative diseases, with over 1500 patient samples. Experimental results demonstrate that our framework consistently outperforms representative state‐of‐the‐art methods. Specifically, our method achieves 66.02% Accuracy and 60.59% Normalized Mutual Information on all categories, and 54.53% Accuracy and 16.67% Normalized Mutual Information on novel categories, surpassing existing state‐of‐the‐art methods. Conclusions Our framework underscores the model's robustness and its practical value in discovering new neurodegenerative disease conditions, highlighting its potential for advancing open‐set diagnostic AI.