White blood cells detection and classification based on regional convolutional neural networks

Area of Science:

• Medical Imaging
• Computational Biology
• Immunology

Background:

• White blood cells (WBCs) are crucial for immune function, and their types and counts offer vital health information.
• Traditional blood cell analysis relies on image processing, which can be limited by cell visibility and overlap.
• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise in overcoming limitations of traditional image analysis.

Purpose of the Study:

• To develop a computer-aided automated system for identifying and locating white blood cell types in blood images.
• To enhance the performance of existing blood test devices using deep learning methodologies.
• To leverage CNNs for more accurate recognition of partially visible or overlapping cells.

Main Methods:

• Utilized Regional Convolutional Neural Networks (R-CNN) for simultaneous classification of multiple cell types within an image.
• Trained and tested CNN architectures (AlexNet, VGG16, GoogLeNet, ResNet50) using combined BCCD and LISC datasets.
• Employed both full learning and transfer learning techniques for model optimization.

Main Results:

• The developed system achieved 100% success in identifying white blood cells.
• ResNet50 with transfer learning demonstrated the highest performance among tested CNN architectures.
• Specific cell type accuracies included Lymphocyte (99.52%), Monocyte (98.40%), Basophil (98.48%), Eosinophil (96.16%), and Neutrophil (95.04%).

Conclusions:

• Deep learning, specifically R-CNN with CNN architectures like ResNet50, significantly improves white blood cell identification accuracy.
• The automated system offers a high-performance solution for analyzing blood cell images, potentially aiding in disease diagnosis.
• Transfer learning with ResNet50 is a highly effective strategy for this automated cell classification task.