Digital Staining of White Blood Cells With Confidence Estimation

Area of Science:

• Medical diagnostics
• Computational pathology
• Digital imaging

Background:

• Chemical staining of blood smears is crucial but costly, time-consuming, and variable.
• Deep generative models offer digital alternatives, but require robust quality assurance for medical safety.

Purpose of the Study:

• To develop a confidence estimation method for digitally stained white blood cells.
• To ensure the quality and trustworthiness of digital staining in critical medical applications.

Main Methods:

• Disentangled latent space representation in Generative Adversarial Networks (GANs) for white blood cell and staining.
• Confidence estimation by noise corruption of representations and quantifying retained information.
• Validation using images from Differential Interference Contrast microscopy and LPIPS metric.

Main Results:

• The proposed confidence estimation method correlates highly with image quality (LPIPS).
• Demonstrated effectiveness in digital staining of white blood cell images from 24 patients.
• The method successfully predicts the quality of generated digital staining outputs.

Conclusions:

• The developed confidence score enhances the reliability of digital staining for medical use.
• This approach ensures trustworthiness in safety-critical medical settings.
• Enables prediction of generated image quality for digital pathology applications.