Enhancing classification of cells procured from bone marrow aspirate smears using generative adversarial networks and...

Area of Science:

• Hematology
• Medical Imaging
• Machine Learning

Background:

• Leukemia is a common cancer in children and adults, necessitating accurate diagnosis through bone marrow cell analysis.
• Manual cell classification from bone marrow smears is labor-intensive and time-consuming.
• Machine learning models require substantial data for effective training in medical diagnosis.

Purpose of the Study:

• To generate synthetic microscopic cell images from bone marrow aspirate smears to augment real-world datasets.
• To enhance the accuracy of automated cell classification for hematologic disease diagnosis.
• To evaluate a novel generative adversarial network (GAN) for synthetic data generation.

Main Methods:

• A three-stage approach was employed, starting with expert-curated bone marrow cell image data.
• A three-network generative adversarial network (GAN) model, C-WGAN-GP, was developed to create synthetic cell images.
• A sequential convolutional neural network (CNN) was utilized for cell classification on both original and synthetic datasets.

Main Results:

• The C-WGAN-GP model generated realistic synthetic cell images, achieving an inception score of 14.52 ± 0.10.
• The sequential CNN model, trained on synthetic data, reached a classification accuracy of 96.98%.
• High accuracy (97.5%), sensitivity (97.1%), and specificity (97%) were achieved for neutrophil classification.

Conclusions:

• The proposed three-network GAN architecture effectively produces more realistic synthetic data compared to existing models.
• Utilizing synthetic data with a sequential CNN model significantly improves cell classification accuracy over using original data alone.