Artificial intelligence differentiates prefibrotic primary myelofibrosis with thrombocytosis from essential thrombocythemia using digitized bone marrow biopsy images

Prefibrotic primary myelofibrosis (prePMF) and essential thrombocythemia (ET) are distinct myeloproliferative neoplasms (MPNs) with overlapping clinical features, often leading to diagnostic uncertainty. We developed an artificial intelligence (AI) framework with human interpretability to distinguish prePMF from ET using digitized hematoxylin and eosin-stained bone marrow biopsy (BMB) slides. Trained on an initial cohort of MPN patients with thrombocytosis, the AI model achieved an AUROC of 0.89 and accuracy of 92.3%. To assess the image features guiding predictions, we generated synthetic images which potentially exaggerate disease-specific morphologies. In a blinded survey, hematopathologists reviewed both real and AI-generated images. While human experts frequently agreed with AI predictions on diagnosis with real images, diagnostic discordance reached up to 88% for AI-generated ET images despite being correctly predicted by AI. We further quantified marrow cellularity and adiposity in the real and generated images, which revealed a higher proportion of fat content in all ET images (42.0%) compared to prePMF (28.9%). These findings suggest that AI can utilize morphological cues distinct from current established diagnostic criteria, such as proportion of adiposity to distinguish types of MPNs. Thus, an AI-assisted diagnostic tool underscores the potential of AI to augment histopathologic evaluation and allow identification of more specific subpopulations of forms of MPNs.