Mobile Menu

Deep learning objectively predicts blood quality

Researchers have used flow cytometry and deep learning to characterise red blood cell lesions in order to objectively predict blood quality.

Storage and Quality

Red blood cells (RBCs) are vital for life-saving blood transfusion, yet there is a worldwide shortage. When in storage, RBCs are subjected to degradation, resulting in oxidative damage, decreased oxygen release capability and membrane deformation. As a result, this can affect in vivo circulation of transfused RBCs. Advancements in preservation and storage has allowed for some countries to store RBCs in blood banks for up to 8 weeks. Nevertheless, during storage, these cells lose membrane integrity, which can disturb capillary blood flow and oxygen delivery. Consequently, this decreases the safety and efficacy of the transfusion.  

Specialists often assess the quality of RBCs using microscopic examination or biochemical and biophysical assays. This process is complex, time-consuming and sometimes destructive to the fragile cells. These methods are also prone to subjective bias. Therefore, improved methods to objectively assess degradation events would enhance quality assessment.

Deep learning

Deep learning has shown promise in detecting biomedically important cell states in images. Therefore, the team hypothesised that a deep-learning-based morphological assessment approach may be a reliable option for RBC quality. In this study, published in PNAS, researchers tested deep-learning methods on RBC images from three independent cohorts in two different countries. They also used imaging flow cytometry (IFC) to assess whether a neural network could be trained to replicate an expert’s judgment in classifying stages of RBC degradation from cell images. This was also to determine whether a neural network might extract subtle features more objectively than humans.

Results from a neural network

The team found that a trained neural network achieved a 76.7% agreement with experts in classifying seven clinically relevant RBC morphologies associated with storage lesions. This was comparable to 82.5% agreement between different experts.

The researchers then removed subjective human annotation from the training step, by training a weakly supervised neural network using only storage duration times. In this approach, the neural network learns the morphological properties of RBCs independently from visual categories defined by experts. Here, the team found that this approach went beyond human visual assessment. The approach discovered the natural progression of RBC deterioration. It matched physiologically relevant physical tests of RBC quality better than manual morphology assessment by experts. This approach also avoided assessment subjectivity.

This study demonstrates the promise of deep learning and IFC in routinely and objectively assessing RBC storage lesions. This automated process will help minimise laboratory sample handling and preparation, and reduce the impact of errors and discrepancies. Experts could also potentially use it to assess morphological changes in other biomedically important progressions such as differentiation and metastasis.

Image credit: Infographic vector created by pch.vector –

More on these topics

Blood / Deep Learning / Neural Networks

Share this article