Deep learning-based morphological assessment of myelodysplastic syndrome on bone marrow smears

Myelodysplastic syndrome (MDS) is a hematological malignancy characterized by the clonal expansion of mutated hematopoietic stem and progenitor cells, resulting in morphological changes (dysplasia) in the bone marrow and cytopenia. Despite advancements in molecular diagnostics, MDS diagnosis often depends on subjective morphological evaluation of bone marrow smears. One objective criterion is cytogenetic analysis. Here, we present a deep learning-based approach to identify morphological abnormalities associated with cytogenetically defined MDS (cMDS) using whole slide images (WSIs) of bone marrow smears. To disentangle shape and textural features in WSIs, we developed a model analyzing segmented binary masks of red blood cells (RBCs) in addition to a conventional image classification model based on color images. Notably, the RBC morphology model achieved high diagnostic accuracy, even after controlling for anemia, uncovering a strong association between RBC shape and distribution-features previously overlooked by human experts-and cMDS. These findings propose RBC morphology as a novel biomarker for MDS and underscore the potential of artificial intelligence to enhance diagnostic precision and objectivity.