Highly-Efficient Differentiation of Reactive Lymphocytes in Peripheral Blood Using Multi-Object Detection Network...

Area of Science:

• Medical Imaging
• Computer Vision
• Hematology

Background:

• Reactive lymphocytes, morphologically transformed leukocytes, are crucial indicators of viral infections.
• Manual detection of these cells is labor-intensive, time-consuming, and requires specialized expertise.
• Computer-assisted diagnosis holds significant promise for improving the efficiency and accuracy of leukocyte analysis.

Purpose of the Study:

• To develop and apply an advanced multi-object detection network for reactive lymphocyte and general leukocyte detection in medical images.
• To enhance the capabilities of deep learning models for practical medical diagnostic scenarios.

Main Methods:

• Implementation of space-to-depth convolution (SPD-Conv) to improve detection of dense, small objects.
• Integration of a dynamic large kernel attention (DLKA) mechanism for better contextual modeling of cells.
• Development of an asymptotic feature pyramid network (AFPN) for robust multi-scale feature fusion.
• Utilized a novel peripheral blood cell dataset for training and validation.

Main Results:

• Achieved a mean Average Precision (mAP50) of 0.918 for reactive lymphocyte detection.
• Attained an mAP50 of 0.907 for the detection of all leukocytes.
• Demonstrated good model interpretability, aiding clinical understanding.
• Provided a new dataset to support future research in peripheral blood cell analysis.

Conclusions:

• The proposed deep learning model represents a significant advancement in the automated detection of reactive lymphocytes.
• The novel network architecture effectively addresses challenges in detecting small, dense objects and fusing multi-scale features.
• This work offers a valuable tool for improving disease diagnosis and provides a foundation for further research in computational hematology.