Deep learning differentiates myeloproliferative neoplasms and predicts JAK2 and CALR mutations from bone marrow smears

Abstract Accurate differentiation of BCR::ABL1-negative myeloproliferative neoplasms (MPNs) based on bone marrow smear morphology remains challenging. While previous artificial intelligence (AI) approaches have primarily focused on bone marrow core biopsies, the diagnostic potential of routine bone marrow smears has remained largely unexplored. Using 460 digitized bone marrow smear whole-slide images, we developed an attention-based multi-instance learning model using features extracted from a self-supervised vision foundation model to classify polycythemia vera (PV), essential thrombocythemia (ET), and primary myelofibrosis (PMF) and to predict JAK2 and CALR mutation status. Our model achieved a receiver-operating-characteristic area-under-the-curve (ROC-AUC) of 0.84 for MPN subtype classification and predicted JAK2 and CALR mutation status in ET (ROC-AUC 0.73 and 0.75) and PMF (ROC-AUC 0.67 and 0.70) based on bone marrow smear images only. Our results provide a proof of-concept for AI-supported bone marrow cytomorphology as a modality complementary to core biopsies in MPN diagnostics and suggest that driver mutations leave traceable morphological signatures in smears. Multimodal integration of microscopic imaging with laboratory and clinical findings may further boost MPN differentiation performance of diagnostic AI models.