Image-based explainable artificial intelligence accurately identifies myelodysplastic neoplasms beyond conventional signs of dysplasia.

Cytomorphological assessment of bone marrow smears (BMS) is essential in the diagnosis of myelodysplastic neoplasms (MDS), yet manual evaluation is prone to inter-observer variability. We trained end-to-end deep learning models to distinguish between MDS, acute myeloid leukemia, and bone marrow donor BMS with high accuracy in internal tests and external validation. Occlusion sensitivity mapping revealed the high importance of nuclear structures beyond canonical dysplasia, demonstrating accurate, interpretable MDS detection without labor-intensive cell-level annotation.