Development of an AI algorithm for the automatic detection and classification of rat bone marrow cells

Area of Science:

• Hematology
• Computational Biology
• Toxicology

Background:

• Morphological classification of bone marrow cells is crucial for pharmaceutical toxicity studies.
• Expertise in light microscopy for cell classification requires extensive training, leading to high labor and time costs.
• Objective and efficient data acquisition is needed for reliable toxicity evaluations.

Purpose of the Study:

• To develop an automated system for detecting and classifying rat bone marrow cells using machine learning.
• To improve the efficiency and objectivity of bone marrow cell analysis in toxicity testing.
• To create a supportive tool for pharmaceutical toxicity evaluation.

Main Methods:

• A machine learning system integrating SSD300 for object detection, VDSR for super-resolution, and EfficientNetV2B0 for classification was developed.
• Whole slide images (WSIs) of rat bone marrow smear specimens were utilized.
• Models were fine-tuned using 720 images for detection and 144,000 cell images for classification of 13 cell types.

Main Results:

• The automated system achieved an average precision of 79% for bone marrow cell detection.
• Morphological classification accuracy reached 98% when compared to expert classifications.
• The system classified approximately 5,000 cells per image in about 80 seconds.

Conclusions:

• The developed algorithm enables highly accurate and rapid automated classification of rat bone marrow cells.
• This system can significantly reduce the time and labor associated with bone marrow cell analysis.
• The algorithm shows potential as a valuable supportive tool for pharmaceutical toxicity evaluations.