Assessing red blood cell deformability from microscopy images using deep learning.

Area of Science:

• Biomedical Engineering
• Hematology
• Artificial Intelligence

Background:

• Red blood cell (RBC) deformability is crucial for oxygen delivery via microvasculature.
• Impaired RBC deformability in disease, aging, or storage affects cell function.
• Current measurement methods are equipment-intensive, time-consuming, and require expertise.

Purpose of the Study:

• To develop a machine learning model for predicting RBC deformability from single-cell images.
• To assess the feasibility of using image-based morphological features for deformability prediction.
• To establish a rapid and accessible method for evaluating RBC deformability.

Main Methods:

• Utilized a microfluidic ratchet device to sort RBCs by deformability.
• Imaged sorted RBCs to extract morphological features.
• Trained a deep learning model to classify RBC deformability based on image features.

Main Results:

• The deep learning model achieved 81 ± 11% accuracy in predicting individual RBC deformability across ten donors.
• Sample deformability scoring using the model showed 10.4 ± 6.8% accuracy compared to the microfluidic ratchet device.
• Demonstrated that deep learning can assess RBC deformability, a property not typically measurable by imaging.

Conclusions:

• Machine learning, specifically deep learning on microscopy images, offers a novel approach to assess RBC deformability.
• This imaging-based method is rapid and utilizes standard microscopy, potentially integrating into routine clinical assessments.
• The developed model provides accurate predictions, overcoming limitations of existing RBC deformability measurement techniques.