White blood cells identification system based on convolutional deep neural learning networks.

Area of Science:

• Medical diagnostics
• Computational biology
• Image analysis

Background:

• Automated white blood cell (WBC) differential counting is crucial for health and disease assessment.
• Current systems rely on multi-stage image analysis, facing limitations with small datasets.
• Deep learning offers potential to enhance WBC identification accuracy.

Purpose of the Study:

• To develop a novel deep learning system for automated WBC identification.
• To address the challenge of classifying limited datasets using deep learning.
• To improve upon existing automated cell morphology analysis methods.

Main Methods:

• Proposed a novel WBC identification system using deep convolutional neural networks.
• Employed two transfer learning methodologies: deep activation features and fine-tuning.
• Developed an end-to-end deep architecture, WBCsNet, trained from scratch.
• Classified a limited, balanced WBC dataset using WBCsNet as a pre-trained network.

Main Results:

• Utilized three public WBC datasets (2551 images) with 5 healthy cell types.
• Achieved an overall system accuracy of 96.1% with the proposed WBCsNet.
• Demonstrated superior performance compared to transfer learning and traditional methods.
• Visualized WBCsNet activations, showing higher responses than pre-trained networks.

Conclusions:

• A novel deep learning-based WBC identification system was successfully developed.
• The proposed WBCsNet demonstrates high performance and potential as a pre-trained network.
• Deep learning effectively addresses challenges in classifying limited medical image datasets.