Morphology-based cytogenetic risk prediction in multiple myeloma from bone marrow smears

Cytogenetic abnormalities determine prognosis and treatment selection in multiple myeloma (MM). However, genetic testing via fluorescence in situ hybridization remains expensive and inaccessible in many clinical settings. Prior studies have established that specific chromosomal alterations correlate with distinct plasma cell morphological features. This suggests that visual analysis of bone marrow aspirate (BMA) smears could enable genetic risk prediction. We developed a computational framework that predicts high-risk cytogenetic status directly from digitized BMA smear images. Our approach employs DinoBloom, an established hematology-specific vision transformer pretrained on bone marrow and peripheral blood cell images, to extract morphological features from individual plasma cells. These embeddings are aggregated through attention-based multiple instance learning, which requires only patient-level labels rather than cell-by-cell annotations. The attention mechanism identifies which cells drive each prediction and links morphological phenotypes to genetic subtypes. Evaluated on a multi-center MM cohort with matched genetic results, the model achieved area under the curve values ranging from 0.76 to 0.85 across multiple cytogenetic markers. Attention maps highlighted plasmablastic features in high-risk predictions and mature plasma cell characteristics in standard-risk cases, consistent with established morphology-genetics correlations. This approach offers a rapid, cost-effective screening tool that could extend genetic risk stratification to resource-limited settings.