Deep learning for bone marrow cell detection and classification on whole-slide images.

Area of Science:

• Hematology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Bone marrow (BM) examination is critical for diagnosing and managing hematologic disorders.
• Traditional BM nucleated differential count (NDC) analysis relies on manual microscopy, which is time-consuming and subjective.
• Automated analysis of whole-slide images (WSIs) for BM NDC faces challenges due to large data volumes and complex cell morphologies.

Purpose of the Study:

• To develop a fully automatic, efficient deep learning framework for BM NDC analysis on WSIs.
• To enable rapid analysis of BM WSIs, potentially replacing traditional manual counting methods.
• To accurately identify a comprehensive range of 16 cell types, including challenging categories like erythroblasts and megakaryocytes.

Main Methods:

• A hierarchical deep learning framework was developed, comprising three stages: ROI localization, patch-based cell identification, and result stitching.
• The framework utilizes a 40x objective magnification for WSI analysis.
• The system was trained and validated on extensive datasets of annotated BM cells.

Main Results:

• The proposed method achieved high recall (0.905 ± 0.078) and accuracy (0.989 ± 0.006) on the primary dataset.
• Analysis of a WSI for BM NDC was completed in just 44 seconds.
• Independent validation demonstrated robust generalizability with recall of 0.842 and accuracy of 0.988.
• The system outperformed existing small-image-based benchmark methods in recall, accuracy, and speed.

Conclusions:

• The developed hierarchical deep learning framework offers an efficient and accurate solution for automated BM NDC analysis using WSIs.
• This AI-driven approach has the potential to significantly improve the speed and consistency of hematologic disorder diagnosis.
• The study represents a significant advancement in applying deep learning to complex histopathological image analysis.