A deep learning model for detection of leukocytes under various interference factors

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate leukocyte detection is vital for diagnosing blood system diseases.
• Current manual methods are time-consuming and require significant expertise.
• Existing automated methods lack full automation or sufficient performance, necessitating better datasets and strategies.

Purpose of the Study:

• To develop an intelligent, high-performance automatic detection strategy for leukocytes using deep learning.
• To establish a comprehensive dataset for leukocyte detection, incorporating clinical interference factors.
• To evaluate and propose an improved ensemble model for enhanced detection accuracy and robustness.

Main Methods:

• Creation of a new dataset with 6273 images (8595 leukocytes) accounting for nine common clinical interference factors.
• Performance evaluation of six mainstream deep learning detection models.
• Development and proposal of a robust ensemble model integrating multiple detection strategies.

Main Results:

• The proposed ensemble model achieved a mean average precision (mAP) of 0.853 and mean average recall (mAR) of 0.922 on the test set.
• Demonstrated robust detection performance even on poor-quality images.
• Achieved a novel 98.84% accuracy in detecting incomplete leukocytes, outperforming existing methods.

Conclusions:

• The deep learning-based ensemble model significantly enhances automated leukocyte detection performance.
• The developed dataset and model offer a valuable resource for advancing research in hematological diagnostics.
• The findings provide a foundation for more efficient and accurate clinical diagnosis of leukocyte disorders.