Biologically-Constrained Multi-Label Classification with Learnable Domain Knowledge

Abstract

Although recent foundation models trained in a self-supervised setting have shown promise in cellular image analysis, they often produce biologically impossible predictions when handling multiple concurrent abnormalities. This is a problem, as the biological information that may be needed for the different clinical-oriented problems is not directly presented in the images. In this study, we present a novel and modular approach to enforce biological constraints in multi-label medical imaging classification. Building on the powerful and rich representations of the DinoBloom hematological foundation model, our method combines learnable constraint matrices with adaptive thresholding, effectively preventing contradictory predictions while maintaining high sensitivity. Extensive experiments on three datasets, two public and one in-house on neutrophil classification, demonstrate significant improvements over different foundation models and the state-of-the-art methods. Through detailed ablation studies and hyperparameter interpretation, we show that our approach successfully captures biological relationships between different abnormalities.