WBC image classification and generative models based on convolutional neural network.

Area of Science:

• Medical Imaging Analysis
• Computational Biology
• Machine Learning in Healthcare

Background:

• Manual white blood cell (WBC) counting is complex and time-consuming.
• Accurate classification of WBCs is crucial for diagnosing immune system conditions.
• Existing computer-aided methods excel at WBC detection but struggle with classification.

Purpose of the Study:

• To develop an accurate and efficient deep learning model for classifying five types of WBCs.
• To create a synthetic dataset of WBC images for educational and research purposes.
• To demonstrate the effectiveness of the proposed model in transfer learning scenarios.

Main Methods:

• Proposed W-Net, a Convolutional Neural Network (CNN)-based architecture for WBC classification.
• Evaluated W-Net on a large-scale dataset of 6562 real microscopic blood images.
• Generated synthetic WBC images using a Generative Adversarial Network (GAN) for data augmentation and sharing.

Main Results:

• W-Net achieved an average classification accuracy of 97% for the five WBC types.
• Outperformed existing state-of-the-art CNN- and Recurrent Neural Network (RNN)-based models.
• Demonstrated successful application of pre-trained W-Net in transfer learning tasks.
• Synthetic images showed high similarity to real images, validated by experts.

Conclusions:

• W-Net provides a highly accurate, CNN-based solution for classifying five WBC types.
• The model effectively addresses challenges like class imbalance and transfer learning.
• The released synthetic dataset and pre-trained model facilitate further research and education in hematology and AI.
• This work advances automated analysis of blood cell images.