Deep Learning Enables High-Precision Classification of Morphology of Stored Red Blood Cells
MetadataShow full item record
About 15 million unit of stored red blood cells are transfused to nearly 5 million patients each year in the United States alone. However, the key properties of red blood cells progressively deteriorate during hypothermic storage, and this reduction in quality contributes to adverse outcomes that have been associated with blood transfusions, including serious infections and multiple organ failures. Recently, red blood cell shape (morphology) has emerged as a novel quantitative marker of the functional quality of stored blood units. Red blood cell morphology is currently measured via manual observation and classification of morphology of only ~100-200 individual cells by a technician, which is a notoriously tedious and highly subjective process. Building on prior work from Dr. Sergey Shevkoplyas’s Blood Microfluidics Lab, we trained an AlexNet Deep Learning CNN with the objective of creating a simple, robust, and automated system for high throughput microscopy that classifies heterogeneous cell morphology. The foundation of this project was successful at predicting morphological classifications with an overall low-resolution accuracy of 97.9% and an overall high-resolution accuracy of 95.3%.