Fine-grained leukocyte classification with deep residual learning for microscopic images

Area of Science:

• Medical image analysis
• Computational biology
• Machine learning

Background:

• Leukocyte classification is crucial in medicine, but traditional machine learning methods struggle with fine-grained categorization of up to 40 white blood cell types from microscopic images.
• Extracting distinctive features and handling complex classification tasks remain challenging for existing techniques like Support Vector Machines (SVM).

Purpose of the Study:

• To develop an advanced deep learning model for accurate fine-grained leukocyte classification from microscopic images.
• To enhance automated cell identification by mimicking expert recognition processes and robust feature extraction.

Main Methods:

• A deep residual neural network (DRNN) classifier was designed, imitating expert cell recognition for robust, automatic feature extraction.
• The DRNN topology was optimized using prior knowledge of white blood cell tests.
• A large dataset of nearly 100,000 labeled leukocytes across 40 categories was created, employing combined training strategies for generalization.

Main Results:

• The DRNN classifier achieved a top-1 accuracy of 77.80% and a top-5 accuracy of 98.75% during training.
• The model demonstrated an average accuracy of approximately 76.84% on the test set for 40 leukocyte categories.
• Experimental results confirmed the feasibility and effectiveness of the proposed fine-grained classification method.

Conclusions:

• The developed deep residual learning-based method offers a robust solution for fine-grained leukocyte classification in microscopic images.
• The approach shows potential for real-world medical applications, assisting physicians in disease diagnosis and significantly reducing manual labor.
• The study highlights the effectiveness of integrating deep learning with medical domain knowledge for improved hematological analysis.