Computational Hematology

Computational haematology applies AI to blood — the most frequently sampled tissue in medicine. Many blood disorders are rare, only a fraction of cells in a sample is altered, and differences between laboratories, stains and microscopes make models brittle.

Together with Carsten Marr at Helmholtz Munich we developed explainable AI for rare blood disorders: attention-based and anomaly-aware multiple instance learning that diagnoses a patient from single-cell images while showing which cells drove the decision, and Percoll density-gradient analysis as an affordable biomarker for hereditary haemolytic anaemias and sickle cell disease. With Siemens and TU Munich we replaced the chemical staining of blood smears with digital staining — generative models that stain, segment and classify white blood cells and estimate how far each synthetic image can be trusted.

The common thread is learning cell-level explanations from patient-level labels and making models robust across sites — principles that carry over to pathology and other rare diseases.