DinoBloom: A Foundation Model for Generalizable Cell Embeddings in Hematology

Abstract: In hematology, computational models offer significant potential to improve diagnostic accuracy, streamline workflows, and reduce the tedious work of analyzing single cells in peripheral blood or bone marrow smears. However, clinical adoption of computational models has been hampered by the lack of generalization due to large batch effects, small dataset sizes, and poor performance in transfer learning from natural images. To address these challenges, we introduce DinoBloom, the first foundation model for single cell images in hematology, utilizing a tailored DINOv2 pipeline. Our model is built upon an extensive collection of 13 diverse, publicly available datasets of peripheral blood and bone marrow smears, the most substantial open-source cohort in hematology so far, comprising over 380,000 white blood cell images. To assess its generalization capability, we evaluate it on an external dataset with a challenging domain shift. We show that our model outperforms existing medical and non-medical vision models in (i) linear probing and k-nearest neighbor evaluations for cell-type classification on blood and bone marrow smears and (ii) weakly supervised multiple instance learning for acute myeloid leukemia subtyping by a large margin. A family of four DinoBloom models (small, base, large, and giant) can be adapted for a wide range of downstream applications, be a strong baseline for classification problems, and facilitate the assessment of batch effects in new datasets. All models are available at github.com/marrlab/DinoBloom.

• The paper introduces DinoBloom, a foundation model for single cell hematology image analysis that enhances classification and diagnostic precision.
• It employs a modified DINOv2 pipeline with vision transformers, fine-tuned on over 380K images from blood and marrow smears to capture detailed cell features.
• Experimental benchmarks show DinoBloom outperforms conventional models in cell-type and AML subtype classification, streamlining diagnostic workflows.

Introducing DinoBloom: A Foundation Model for Single Cell Image Analysis in Hematology

Overview of DinoBloom

In the field of hematology, there has been a persistent challenge in streamlining the analysis of peripheral blood and bone marrow smears, pivotal for diagnosing blood-related diseases. The novel introduction of DinoBloom, a series of foundation models created for the analysis of single cell images within this domain, marks a significant stride toward overcoming these hurdles. Derived from an extensive assembly of datasets, encompassing over 380,000 white blood cell images, DinoBloom is poised to enhance diagnostic processes, alleviate labor-intensive workflows, and augment the precision of medical evaluations in hematology.

Methodological Approach

DinoBloom leverages a tailored DINOv2 pipeline for its development, adapted specifically to the nuances of single cell images in hematology. The model diverges from the conventional approach by eliminating the global-local crop loss, a modification that has shown to improve performance in learning representations of single-cell imagery significantly. Four variations of the DinoBloom model, ranging from small to giant, were fine-tuned on this diversified dataset collection, employing vision transformers (ViT) of different sizes. The unique composition of the training data, inclusive of both peripheral blood and bone marrow smears, facilitated a comprehensive learning scope, enabling the models to capture a wide array of visual features inherent to different cell types and states.

Experimental Evaluations and Findings

The evaluation of DinoBloom's performance encompassed a series of rigorous benchmarks. These included linear probing and k-nearest neighbor assessments for cell-type classification across blood and bone marrow smears and a weakly supervised multiple instance learning setup for the subtyping of acute myeloid leukemia. The findings illustrated that DinoBloom models significantly outperformed both medical and non-medical vision models across these tasks, showcasing remarkable generalization capabilities, particularly in the presence of challenging domain shifts.

• Performance Metrics:
  • DinoBloom models reported leading performance figures in linear probe and k-NN evaluations for cell-type classification.
  • In the domain of AML subtype classification through multiple instance learning, DinoBloom demonstrated superior efficacy, outpacing other models by significant margins.

Implications and Potential Applications

The practical implications of DinoBloom are twofold, impacting both theoretical understanding and clinical applications in hematology. Theoretically, the development of DinoBloom enriches the foundation model landscape, introducing a resource uniquely tailored to hematology's intricacies. Clinically, DinoBloom presents a groundbreaking tool for refining diagnostic accuracy, offering a robust baseline for classification problems, and streamlining the cumbersome tasks of manually analyzing single cells in blood smears.

• Future Orientations:
  • The adaptability of DinoBloom for an expansive range of downstream applications opens avenues for further research and development in automated diagnostics.
  • The potential for DinoBloom's cell embeddings to characterize disease profiles and facilitate interpretability positions the model as a pivotal tool for advancing clinical decision-making processes.

Conclusion

DinoBloom stands as the first extensive, self-supervised trained model explicitly designed for single cell hematology image analysis, equipped with the unparalleled ability to navigate the challenges of dataset variance, batch effects, and the exigencies of weakly-supervised learning settings. By making DinoBloom models and their training specifications openly accessible, the research endeavors to catalyze further innovation and collaborative efforts within the community, fostering advancements that could redefine diagnostic methodologies in hematology.