Segmentation, feature extraction and classification of leukocytes leveraging neural networks, a comparative study

Area of Science:

• Medical Image Analysis
• Computational Biology
• Machine Learning

Background:

• Manual leukocyte differentiation from blood smears is laborious and prone to human error.
• Existing automated methods lack comprehensive comparative studies on segmentation, feature extraction, and classification.

Purpose of the Study:

• To compare traditional machine learning with various deep learning models for leukocyte segmentation and classification.
• To evaluate the performance of different deep learning architectures for feature extraction and cell classification.

Main Methods:

• Cell segmentation using K-means clustering (traditional ML) versus U-Net, U-Net+ResNet18, U-Net+ResNet34 (deep learning).
• Feature extraction and classification using AlexNet, VGG16, and ResNet18 (deep learning).
• Validation on CellaVision, BCCD, ALL-IDB2, and PCB-HBC datasets.

Main Results:

• Deep learning models achieved higher segmentation accuracy (e.g., 99.17%) than K-means (94.36%) on the CellaVision dataset.
• Deep learning classification models, particularly ResNet18, reached 100% accuracy on the CellaVision dataset.
• Demonstrated high classification accuracies (100% and 98.49%) on ALL-IDB2 and PCB-HBC datasets.

Conclusions:

• Deep learning models significantly enhance accuracy and efficiency in automated leukocyte classification compared to traditional machine learning.
• Increasing neural network depth improves leukocyte classification performance.
• The validated deep learning models show strong potential for clinical applications in hematology.