Deep learning approach to peripheral leukocyte recognition.

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Manual microscopic examination of peripheral blood leukocytes is crucial for disease diagnosis but is time-consuming, labor-intensive, and subjective.
• Traditional automated systems using feature engineering struggle with robustness and require successful cell segmentation.
• Convolutional neural network classification pipelines automate feature extraction but face challenges in multi-object recognition.

Purpose of the Study:

• To investigate the application of object detection approaches for automated peripheral leukocyte recognition.
• To enhance the accuracy and efficiency of leukocyte classification compared to traditional methods.
• To explore key factors influencing object detection performance in this context.

Main Methods:

• Peripheral leukocyte recognition was framed as an object detection task.
• Two object detection models, Single Shot Multibox Detector (SSD) and an improved You Only Look Once (YOLO) version, were applied.
• Models were trained on 14,700 annotated images and evaluated on test sets of 1,120 annotated images and 7,868 single-object images across 11 leukocyte categories.

Main Results:

• The developed object detection models achieved a best mean average precision (mAP) of 93.10%.
• A mean accuracy of 90.09% was obtained for peripheral leukocyte recognition.
• The models demonstrated efficient inference times of 53 ms per image on a NVIDIA GTX1080Ti GPU.

Conclusions:

• Object detection models, specifically SSD and improved YOLO, offer a robust and efficient solution for automated peripheral leukocyte recognition.
• These AI-driven approaches significantly improve upon the limitations of manual analysis and traditional automated systems.
• The high accuracy and speed achieved pave the way for enhanced clinical diagnostics and disease control.