Truace
Subscribe

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 know…

By the Truvace desk1 min read1 sourcePeer-reviewedTRV-2026-2187· cite & verify
Learning global and local representations for generalized category discovery in neurodegenerative diseases
Dance training is superior to repetitive physical exercise in inducing brain plasticity in the elderly by Kathrin Rehfeld, Angie Lüders, Anita Hökelmann, Volkmar Lessmann, Joern Kaufmann, Tanja Brigadski, Patrick Müller, Notger G. Müller. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

In brief

Researchers built a framework for generalized category discovery in neurodegenerative diseases using MRI. It uses labeled data from known disease categories to automatically group unlabeled scans into new, previously unseen categories, combining semi-supervised contrastive learning with global and local representation learning.

This matters because AI classifiers that work well on fixed, known disease labels tend to degrade in realistic open-set settings where rare or novel subtypes appear. The authors describe their result as potential for open-set diagnostic AI, but the evidence is a benchmark evaluation, not a clinical deployment.

Main points

  1. The authors describe it as the first generalized category discovery framework for neurodegenerative diseases using MRI.
  2. Globally, the method aligns similarity relationships between weakly augmented views of the same instance and other embeddings in a support queue.
  3. Locally, neighborhood embeddings from the support queue serve as positive pairs for cross-instance contrastive learning.
  4. The benchmark included six neurodegenerative diseases and over 1500 patient samples.

The gain

A new MRI-based category discovery framework outperformed state-of-the-art methods at identifying novel neurodegenerative disease categories in unlabeled data, using labeled data from known categories.

The rundown

The study addresses the drop in AI performance when moving from closed-set classification to open-set settings. In the open-set case, models must discover rare or novel disease subtypes from unlabeled data not seen during training.

The method trains one unified objective applicable to both labeled and unlabeled data, narrowing the representation gap among semantically similar instances. The aim is finer-grained discrimination between disease categories.

On all categories the method reached 66.02% accuracy and 60.59% normalized mutual information. On novel categories it reached 54.53% accuracy and 16.67% normalized mutual information, which the authors report as surpassing existing methods.

Sources

  1. Peer-reviewedMedical Physics2026-10-22

The debate