Machine-based detection and classification for bone marrow aspirate differential counts: initial development focusing...

Area of Science:

• Hematology
• Digital Pathology
• Machine Learning

Background:

• Manual bone marrow differential cell counts (DCCs) are crucial for diagnosing hematologic disorders but are labor-intensive and prone to bias.
• Existing automated methods are unreliable due to the complexity of bone marrow specimens.
• A lack of comprehensive training datasets hinders the development of automated DCC solutions.

Purpose of the Study:

• To develop and validate a machine learning algorithm for automated detection and classification of bone marrow cells.
• To create and utilize annotated datasets for training artificial intelligence models for DCCs.
• To address the need for objective and efficient methods in hematologic disorder diagnostics.

Main Methods:

• Development of a web-based system for annotating and managing digital pathology images of bone marrow aspirate (BMA) smears.
• Manual annotation of over 10,000 BMA cells across all standard clinical DCC classes.
• Implementation of a two-stage machine learning approach for cell detection and classification.

Main Results:

• The developed algorithm achieved high accuracy in cell detection (0.959 ± 0.008 precision-recall AUC) and classification (0.982 ± 0.03 ROC AUC) on nonneoplastic samples.
• The algorithm demonstrated similar performance on limited acute myeloid leukemia and multiple myeloma sample sets.
• The study successfully created essential training datasets for BMA cellular constituents.

Conclusions:

• The machine learning algorithms show promising results for automating bone marrow differential cell counts.
• This work represents a significant initial step towards a reliable and objective automated DCC system.
• Further clinical validation is needed, but the technology has the potential to significantly impact disease diagnosis and prognosis.