Abstract
Discovering novel biomedical concepts—such as rare cell types or emerging disease variants—from unlabeled imaging data is a critical challenge in medical AI. Standard Generalized Novel Class Discovery (GNCD) methods often fail due to non-i.i.d. data distributions and severe class imbalance. We present a framework that models data as probabilistic von Mises-Fisher (vMF) distributions on a unit hypersphere, constrained by uniformly distributed a priori proxies. This geometric approach ensures clear separation between known base classes and novel discoveries. Furthermore, we introduce a spectral-based method for efficient class number estimation. Our framework demonstrates superior performance across four diverse biomedical tasks, achieving significant gains in novel class accuracy and computational efficiency.
Introduction
In biomedical research, identifying "unseen" categories within large, unlabeled datasets is essential for scientific discovery. However, Generalized Novel Class Discovery (GNCD) is complicated by two primary factors: non-i.i.d. distributions arising from varying imaging protocols and extreme class imbalance where rare phenotypes are often overlooked. Existing models frequently suffer from representation collapse, where novel classes are misidentified as common known categories.
Methods
The proposed approach leverages geometry-constrained probabilistic modeling to structure the feature space. Instead of deterministic point embeddings, images are mapped to von Mises-Fisher (vMF) distributions on a unit hypersphere.
Key methodological innovations include:
- Uniform Proxies A Priori: The model initializes equidistant "anchors" across the hypersphere using energy minimization, providing a structured map that prevents base classes from dominating the space.
- Boundness and Dispersion: Known classes are constrained to specific territories around these anchors, while high-uncertainty instances are pushed into the "open space" to form distinct novel clusters.
- Spectral Class Estimation: To avoid manual specification of the number of novel classes, a spectral graph theory approach analyzes the eigenvalues of the graph Laplacian, enabling automatic and 10x faster estimation than previous recursive methods.
Results
The framework was evaluated on datasets for Pneumonia, Cell Nuclei, Skin Lesions, and Diabetic Retinopathy. The method consistently outperformed state-of-the-art models like GCD and DCCL.
In the Pneumonia task, the model achieved 66.67% accuracy for novel classes, a significant improvement over the 60.32% achieved by the best competitor. For Cell Nuclei discovery, the F1-score for novel classes improved by nearly 7%. Qualitative 3D visualizations confirmed that the geometric constraints lead to cleaner, more separated clusters compared to standard methods.
Conclusion
By integrating geometric constraints with probabilistic modeling, this framework provides a robust solution for discovering novel biomedical concepts in challenging, real-world data environments. Future work will explore the sensitivity of performance to proxy density and the extension of this geometric approach to multi-modal data integration.
Summary updated