DinoBloom: A Foundation Model for Hematology Cell Embeddings

DinoBloom exemplifies the potential of foundation models in hematology, focusing on cellular image analysis for diagnosing blood and bone marrow diseases. Tailored from DINOv2 pipeline, DinoBloom is trained on a large open-source collection of cell images, serving as a formidable tool for computational diagnosis.

Distinctive findings of the research:

• DinoBloom outshines both medical and non-medical models in classification accuracy and multiple instance learning challenges.
• The model is a robust baseline for classification problems and useful for assessing batch effects.
• Four model sizes (small, base, large, giant) cater to different computational constraints and applications.

DinoBloom’s innovation signifies a leap in computational tools for hematology, offering a pathway to enhanced precision and efficiency in medical diagnostics. Its generalizability could be instrumental in molding future research and clinical protocols. Read more