Quantitative analysis of blood cells from microscopic images using convolutional neural network

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence in Healthcare

Background:

• Traditional microscopic blood cell counting is labor-intensive and error-prone.
• Existing methods lack detailed morphological analysis (shape, size) crucial for disease diagnosis.
• Accurate blood cell quantification and characterization are vital for clinical decision-making.

Purpose of the Study:

• To develop and evaluate a deep learning-based system for automatic blood cell classification and quantitative analysis.
• To improve the accuracy and efficiency of blood cell counting and morphological assessment.
• To provide a comprehensive analysis of blood cells, including type, count, and morphological parameters.

Main Methods:

• Utilized the YOLOv2 deep learning model for object detection and segmentation of blood cells.
• Trained the model on 1560 images with 2703 labeled blood cells.
• Tested the model on 26 images, analyzing red blood cells, platelets, and various white blood cells (basophils, eosinophils, lymphocytes, monocytes, neutrophils).

Main Results:

• Achieved an average accuracy of 80.6% and precision of 88.4% for blood cell detection and segmentation.
• Obtained high mean accuracy in quantitative analysis: 92.96% for area, 91.96% for aspect ratio, 88.736% for diameter, and 92.7% for cell counting.
• The system successfully classified seven types of blood cells and quantified their morphological parameters.

Conclusions:

• The proposed deep learning approach offers an efficient and accurate alternative to manual blood cell analysis.
• The YOLOv2 model demonstrates significant potential for automated clinical hematology diagnostics.
• This method provides valuable quantitative morphological data, enhancing diagnostic capabilities for blood-related disorders.