Extraction of Novel Features and Diagnosis Prediction in Myelodysplastic Neoplasm Using a Weakly Supervised Artificial Intelligence Model Based on Normal Megakaryocytes.

Abstract

Evaluation of bone marrow pathology can be challenging for nonspecialist pathologists due to its morphological complexities. Despite advances in artificial intelligence for other organ systems, research in bone marrow biopsy field remains limited. This study presents an artificial intelligence model developed to classify myeloid diseases based on morphologically normal megakaryocytes in hematoxylin-eosin-stained bone marrow biopsy specimens. The model integrates two deep learning components: one for detecting bone marrow regions, and the other for identifying megakaryocytes, along with an XGBoost-based classifier that leverages extracted features to differentiate between normal cases, myelodysplastic neoplasm, and immune thrombocytopenic purpura. The model achieved exceptional accuracy, with area under the curve values of 0.9996 (bone marrow detection) and 0.9997 (megakaryocyte detection). For disease classification, myelodysplastic neoplasm versus normal performed well, with an area under the curve of 0.879. Feature analysis revealed that the percentage of megakaryocyte among all cells and the number of adjacent megakaryocytes within various distances were significantly correlated with disease prediction. This study introduces the first artificial intelligence model capable of classifying myelodysplastic neoplasm versus normal based on normal megakaryocyte morphology. This model demonstrates potential not only for diagnostic assistance but also for uncovering novel histological features.