Associating the phenotypic expression of platelets with disease type through image-based single-cell profiling

Platelets are central to hemostasis and thrombosis and contribute to diverse diseases, including cardiovascular disorders, infections, and cancer. Although platelet activation and dysfunction in disease have been studied extensively, disease-associated platelet phenotypes remain poorly defined, largely because phenotypic changes are subtle, transient, and difficult to capture at single-cell resolution.

We acquired bright-field images of numerous circulating platelets in whole-blood samples anticoagulated with 3.2% sodium citrate that were collected from patients (n = 65) and healthy volunteers (n = 10) at the University of Tokyo Hospital, using optofluidic imaging with an optical frequency-division-multiplexed (FDM) microscope integrated with a microfluidic chip. Convolutional neural network (CNN) models were trained to classify platelet phenotypes across disease categories. Feature-importance analysis was performed to identify image-derived parameters driving the deep-learning-based characterization.

We developed three CNN models to classify platelet images by disease category, achieving accuracies of up to 81.3%, indicating measurable associations between platelet phenotypic expression and disease type. Longitudinal analysis of platelet images further showed that the CNN models could predict thrombotic progression up to 7 days before clinical thrombus detection. Feature-importance analysis highlighted texture-related descriptors as the dominant contributors, accounting for 54.5% of overall importance.

Image-based analysis of circulating platelets can capture disease-associated phenotypic signatures and may complement existing clinical workflows for prediagnosis and early assessment of thrombotic risk.