Recognition of peripheral blood cell images using convolutional neural networks

Area of Science:

• Medical Imaging
• Computational Biology
• Hematology

Background:

• Morphological analysis is crucial for diagnosing over 80% of hematological diseases.
• Differentiating blood cell types requires significant expertise.
• Automated classification systems can aid in accurate diagnosis.

Purpose of the Study:

• To develop an automated system for classifying eight groups of peripheral blood cells.
• To achieve high accuracy in cell classification using transfer learning and convolutional neural networks.
• To eliminate the need for image segmentation and automatic feature extraction.

Main Methods:

• A dataset of 17,092 peripheral blood cell images was utilized.
• Two convolutional neural network architectures, Vgg-16 and Inceptionv3, were employed.
• Models were trained using feature extraction with a support vector machine and through fine-tuning.

Main Results:

• Feature extraction with Vgg-16 and Inceptionv3 yielded accuracies of 86% and 90%.
• Fine-tuning the networks resulted in higher accuracies of 96% (Vgg-16) and 95% (Inceptionv3).
• The best overall classification accuracy achieved was 96.2%.

Conclusions:

• A convolutional neural network-based classification scheme was successfully developed.
• Fine-tuning state-of-the-art architectures with clinical data produced a robust end-to-end classifier.
• The system demonstrated excellent precision, sensitivity, and specificity for blood cell classification.