Artificial intelligence: A machine learning model for disease assessment in MPN

Myeloproliferative neoplasms (MPN) are a group of diseases characterized by clonal proliferation of bone marrow stem cells, resulting in increased platelets, red blood cells, or white blood cells. MPN can be subtyped into myelofibrosis, essential thrombocythemia (ET), and polycythemia vera (PV). The differentiation of these subtypes and their reactive mimics is important in the development of individual treatment plans targeting varying presentations of the disease. Subtype differentiation requires morphological assessment of bone marrow trephines (BMT); however, morphological assessment is unreliable due to its subjective nature and reliance on poorly reproducible criteria. Also, MPN treatment is further complicated by overlapping features across subtypes, especially in earlier disease stages.

To improve the assessment of MPN and its subtypes, a machine learning approach was developed by Sirinukunwattana et al. to automate the identification, quantitative analysis, and abstract representation of megakaryocyte features from BMT samples. The MPN Hub is pleased to summarize this model below.

Machine learning model

Overall,[MM1]  131 samples were selected from patients with myelofibrosis, ET, and PV with reactive/non-neoplastic samples sourced from patients with either an established or new diagnosis, satisfying the diagnostic criteria of the latest World Health Organization (WHO) classification.

Step 1: Megakaryocyte identification and delineation

• Detection and delineation of megakaryocytes were conducted from samples through an automated deep-learning approach.
• Image segmentation was required to partition images into regions and determine boundaries of megakaryocyte cells, separating cell areas from the background microenvironment.

Step 2: Megakaryocyte library

• A library of megakaryocytes was developed from an initial set of 2,427 manually annotated megakaryocytes.
• This library was checked by a hematopathologist using a computer-assisted annotation tool to highlight individual cells and prompt acceptance or rejection.
• This was used to identify a set of cytomorphological subtypes for the program to learn.

Step 3: Megakaryocyte phenotyping

• To train the program to phenotype megakaryocytes, a sample of pre-validated cells was used.
• A hematopathologist manually validated the artificial intelligence (AI) decisions to ensure accuracy.
• Cluster analysis was carried out to group cytomorphological similar megakaryocytes.
• The learned subgroups were organized into nine distinct groups using latent representation vectors and Markov clustering.

Step 4: Bone marrow description

• Topographic distribution of the marrow space was mapped and represented to describe megakaryocytes in the samples and create a radar plot for direct comparison.
• To further enhance sample visualization, a principal component analysis was used in an abstract 2-dimensional space to allow a single sample to be indexed into a reference cohort.
• To visualize changes over time, changes in megakaryocyte phenotype were observed as a shift in feature location on multivariable plots (where sequential samples were available).

Step 5: Cohort indexing

• To determine the applicability of cohort indexing, sequential BMT samples were analyzed, including patients with stable ET and PV who had undergone multiple bone marrow biopsies.
  • Plots had remained consistent, highlighting the stability of megakaryocyte samples.
  • In an alternative sample in the MPN subtype had developed, a significant difference in plots was observed.[MM2] 

Model overview

An overview of the machine learning model is shown in Figure 1.

Figure 1. Overview of MLM for the assessment of MPN* 

MLM, machine learning model; MPN, myeloproliferative neoplasm.
*Adapted from Sirinukunwattana, et al.¹

 Applications of AI

Potential benefits of an AI-assisted tool in this indication include,

• faster time to diagnosis and classification, in advance of a formal pathology report;
• simple visual representations of disease features to aid in sample assessment and monitoring disease progression;
• an alternative option when hematopathologist access is limited;
• may assist non-expert clinicians in diagnostic or clinical decision-making; and
• allows for remote diagnosis in areas with limited resources or in lower-income countries.

Conclusion

The machine learning model demonstrated feasibility as a tool to aid in the diagnosis, classification, and monitoring of MPN. A reliable and accurate tool for this indication offers several potential benefits, including prompt diagnosis, while aiding decision-making in resource-limited areas.

However, the current WHO classification system for MPN includes other features, such as marrow cellularity, lineage maturation, degree of fibrosis, and blast cell estimation. Therefore, to provide a comprehensive and viable tool for use in real-world practice, more learned features would be required; this would involve conducting a more in-depth analysis of the bone marrow microenvironment.

To find out more about the role of AI in the future of MPN bone marrow histopathology, watch this discussion featuring Daniel Royston and our steering committee members.