LEMON: a foundation model for nuclear morphology in Computational Pathology

A research team from Institut Curie, Mines Paris PSL, and INSERM has introduced LEMON (Learning Embeddings from Morphology Of Nuclei), a breakthrough foundation model designed specifically for computational pathology. Unlike previous approaches that focused on patch or whole-slide image analysis, LEMON operates at the single-cell level—a crucial but previously underexplored scale for understanding cellular phenotypes in cancer. The model was trained using self-supervised learning on millions of cell images spanning diverse tissues and cancer types, allowing it to learn robust and versatile morphological representations without requiring extensive manual labeling.

LEMON addresses a significant gap in computational pathology by providing scalable tools for single-cell image representation learning. The researchers evaluated the model across five benchmark datasets covering various prediction tasks, demonstrating strong performance that suggests LEMON could become a new paradigm for cell-level analysis. By making model weights publicly available, the team enables researchers worldwide to apply these representations to large-scale single-cell analyses, potentially accelerating discoveries in cancer biology and precision medicine. The model's ability to capture subtle morphological variations at the cellular level could lead to more accurate cancer subtyping, treatment response prediction, and biomarker discovery.