Artificial Intelligence Enhances Diagnostic Precision in Myeloproliferative Neoplasms

Precise distinction between prefibrotic primary myelofibrosis (prePMF) with thrombocytosis and essential thrombocythemia (ET) has been a longstanding challenge in hematology, as these diseases share similar clinical and histopathologic features. They are both grouped as Philadelphia chromosome–negative myeloproliferative neoplasms (MPNs), but they have different prognoses, progression risks, and management strategies. Traditionally, interpretation of bone marrow biopsy has been subjective, and diagnostic discordance among expert pathologists is common, highlighting a significant unmet need for more objective, reproducible approaches in clinical practice.¹

Distinguishing prePMF From ET

Histopathological standards defined by the World Health Organization are primarily based on the morphology of megakaryocytes and the extent of fibrosis in hematoxylin and eosin (H&E)–stained bone marrow biopsy specimens. However, these evaluations are, to some extent, subjective and may vary, particularly in the early stages of the disease when the morphological differences are slight. Data from some studies have demonstrated the high variability in histopathologic diagnosis, to the extent that error rates can lead to misclassification of prePMF as ET, thereby affecting clinical decision-making and patient prognosis.²

Artificial intelligence (AI) and machine learning (ML), which are rapidly advancing, can be seen as valuable adjuncts for pathologists by revealing morphological details that the human eye can miss. These methods are capable of enhancing accuracy, limiting the fluctuations of results, and offering instant diagnostic assistance, which is extremely necessary for rare hematologic cancers whose histopathological features pose a challenge even to highly skilled hematopathologists.¹

Development of an AI Framework for Bone Marrow Analysis

According to data from a recently published study in Leukemia, investigators evaluated an AI-based framework designed to differentiate prePMF with thrombocytosis from ET using digitized bone marrow biopsy images. This AI model was trained using whole-slide H&E images from patients diagnosed with prePMF or ET at institutional sites in Italy and the United States. Notably, the algorithm achieved high accuracy in distinguishing between the 2 conditions, with an overall accuracy of 92.3% and an area under the receiver operating characteristic curve (AUROC) of 0.89, demonstrating robust performance in a blinded validation cohort.¹

The study authors noted that the model’s performance metrics (eg, high specificity and strong accuracy) were very impressive given the challenges of distinguishing these MPN subtypes. Traditional diagnostic methods rely on clinicians’ interpretation, which varies with experience and is not always reproducible across observers.¹

Meanwhile, the AI framework consistently applied the learned image features from the training and validation sets, a strong indication of its usefulness as an objective diagnostic aid for human diagnosticians.

Clinical Implications and Future Directions

The successful use of AI in this setting can have significant implications for clinical practice. As a reproducible, high-performance adjunctive tool, clinicians may be able to overcome the barriers that have long existed in MPN diagnosis, thereby reducing interobserver variability and speeding time to an accurate diagnosis. Consequently, this could lead to better risk stratification, more precise therapy selection, and ultimately, better patient outcomes.¹

Still, incorporating AI into daily pathology practice will require thorough prospective clinical trials, integration with patient data, and clear guidelines on how AI-based outputs should be used for final diagnosis. AI is not to be considered as a substitute for, but rather a complement to, the expert clinical judgment. A close partnership among computer scientists, hematopathologists, and clinicians will be crucial if we are to harness the full potential of these groundbreaking discoveries and bring them into clinical application.¹

Conclusion

The application of AI to digitized bone marrow biopsy images represents a significant advancement in the diagnostic evaluation of myeloproliferative neoplasms. With demonstrated high accuracy and interpretive capability, AI has the potential to enhance diagnostic precision and support more individualized patient care in hematology. As these technologies evolve, they may become invaluable components of integrated diagnostic workflows that bridge computational power and clinical expertise.¹

REFERENCES

1. Srisuwananukorn A, Krull JE, Ma Q, Zhang P, Pearson AT, Hoffman R. Applications of artificial intelligence to myeloproliferative neoplasms: a narrative review. Expert Rev Hematol. 2024;17(10):669-677. doi:10.1080/17474086.2024.2389997

2. Barbui T, Thiele J, Vannucchi AM, Tefferi A. Problems and pitfalls regarding WHO-defined diagnosis of early/prefibrotic primary myelofibrosis versus essential thrombocythemia. Leukemia. 2013;27(10):1953-1958. doi:10.1038/leu.2013.74