Deep learning reveals direct correlation between red blood cell morphology and deformability

Abstract

The deformability of red blood cells (RBCs) is a critical factor in understanding cardiovascular diseases and serves as a key determinant of RBC lifespan. Abnormal RBC deformability can lead to impaired blood flow and an increased risk of thrombosis. Although various methods exist to measure RBC deformability, they often involve complex procedures and require expensive equipment. In both our experiments and existing literature, it has been observed that RBCs with varying deformabilities exhibit significant differences in their morphological characteristics. To further investigate this relationship, we proposed a deep learning-based approach to explore the correlation between RBC deformability and their morphological features. We utilized a dielectrophoretic microfluidic method to assess deformability and capture microscopic images of RBCs. These images were then categorized according to the deformability of the cells. A convolutional neural network model was trained on this dataset for cell classification. Additionally, we validated the efficacy of deep learning methods for image dataset augmentation. The proposed model achieved an accuracy rate exceeding 90%, with an average accuracy of 82% in practical blood tests. Our findings reveal a strong correlation between RBC morphological characteristics and deformability, suggesting that cell images can reliably indicate deformability. This approach has the potential to significantly simplify the study of RBC mechanical properties, with important implications for RBC classification, theoretical research, disease screening, etc.