DeepHeme, a high-performance, generalizable deep ensemble for bone marrow morphometry and hematologic diagnosis.

Area of Science:

• Computational pathology
• Artificial intelligence in hematology
• Medical diagnostics

Background:

• Bone marrow aspirate (BMA) cytomorphology is crucial for diagnosing hematological disorders but is complex and error-prone.
• Current deep learning models for BMA analysis lack expert-level accuracy and generalizability.

Purpose of the Study:

• To develop and validate a deep learning model for accurate and generalizable bone marrow cell classification.
• To achieve expert-level performance in automated hematopathology slide analysis.

Main Methods:

• Developed DeepHeme, a snapshot ensemble deep learning classifier using a curated dataset of 30,394 bone marrow images.
• Trained and tested DeepHeme on data from the University of California, San Francisco, and validated on an independent dataset from Memorial Sloan Kettering Cancer Center.
• Compared DeepHeme's cell classification performance against three human hematopathology experts.

Main Results:

• DeepHeme achieved higher accuracy than previous models and classified more cell types.
• External validation demonstrated robust generalizability across different datasets and WSI systems.
• DeepHeme's diagnostic performance was comparable to or exceeded that of human experts in individual cell classifications.

Conclusions:

• DeepHeme represents a significant advancement in automated hematopathology slide analysis.
• Accurate and generalizable AI-driven cell classification facilitates the development of predictive markers.
• This technology has the potential to improve diagnostic efficiency and accuracy in hematological disorder workup.