Artificial Intelligence-Based Segmentation of Megakaryocytes in Bone Marrow Biopsies
Telling the subtypes of myeloproliferative neoplasms apart is a demanding task in routine pathology. The Department of Pathology at Ankara University Faculty of Medicine and the Simplex Artificial Intelligence Team developed models that detect and classify megakaryocytes in bone marrow biopsies with artificial intelligence and carry this information into diagnosis.
The clinical problem: separating three look-alike diseases
Polycythemia vera (PV), essential thrombocythemia (ET) and primary myelofibrosis (PMF) are the three classical subtypes of BCR::ABL1-negative myeloproliferative neoplasms (MPNs). Although they share overlapping clinicopathological and molecular features, they differ in complications and disease progression, which makes them hard to distinguish in routine practice.
Artificial intelligence can strengthen conventional morphological assessment through quantitative image analysis. The study aimed to develop machine learning models that help pathologists diagnose MPNs and differentiate their subtypes by identifying and analyzing megakaryocyte features.
Method: digital slides, expert labels, AI
Hematoxylin and eosin-stained bone marrow biopsies from 27 PV, 37 ET and 17 prefibrotic PMF cases diagnosed between January 2012 and September 2024, together with 26 non-neoplastic biopsies from patients without hematolymphoid disease or prior chemotherapy, were reviewed retrospectively and digitized with a 3DHISTECH scanner.
In the digital images, megakaryocyte nuclei (normal, bare, separated, bulbous, hyperchromatic, hyperlobated, and monolobated/
Results
The segmentation model detected megakaryocytes with an F1 score of 87.6%. The classification model distinguished dysplastic from normal megakaryocytes with 96.74% accuracy and predicted nuclear morphologic classes with 92.53% accuracy.
Diagnostic models differentiated MPNs from non-neoplastic bone marrow with an average accuracy of 94%, reaching 100% in the best-performing model. For subtype differentiation, accuracy was 81.48% with morphological features alone and rose to 92.59% once clinical and laboratory data were added.
Conclusion
AI-based models showed strong performance in megakaryocyte segmentation and classification and in the diagnosis and subtype differentiation of MPNs. These approaches may support pathologists by automating workflows, accelerating diagnostic processes and providing objective results. The findings point to potential clinical applicability and adaptability to other hematologic malignancies that require quantitative histologic evaluation.