Developing Interpretable Hematology Deep Learning Models For Linking Morphology To Molecular Features In Multiple Myeloma - Abstract Despite therapeutic advances, multiple myeloma (MM) remains an incurable malignancy with heterogeneous outcomes. While molecular features like cytogenetic abnormalities and minimal residual disease status provide important prognostic information, these tests are often unavailable or delayed. In contrast, bone marrow histology slides are routinely collected and contain rich visual information about disease biology. Deep learning models have shown promise in analyzing histology across many cancers, but their application to hematologic malignancies remains limited. Recent advances in foundation models - deep learning systems pre-trained on vast histology datasets - offer potential solutions, but have been developed primarily for solid tumors. We hypothesize that bone marrow histology contains morphological features predictive of molecular characteristics, and that deep learning models can be refined to extract this information. First, we will evaluate and optimize contemporary foundation models for hematologic histology using a dataset of of over 2000 bone marrow specimens with matched molecular and clinical annotations. Second, we will apply deep learning models to identify morphological features in myeloma (using over 200 multiple myeloma cases) that correlate with established prognostic markers, including cytogenetic abnormalities and minimal residual disease status. Through novel explainability methods, we will characterize these morphological patterns to provide biological insights. This work will develop interpretable artificial intelligence tools for hematology while advancing our understanding of myeloma biology.
