Background Neutrophil extracellular trap (NET) formation is shown to have proinflammatory effects. Despite its growing relevance, standardized quantification of NETosis remains challenging, as neutrophil activation occurs in a continuous and morphologically heterogenous spectrum that is difficult to analyze objectively. Automated image-based analysis could provide an approach to quantify neutrophil activation and identify patients with increased inflammatory activity or risk of disease progression.
Aim This study aims to develop a deep learning algorithm capable of detecting, classifying and quantifying different stages of neutrophil activation in health and disease with the goal of establishing NET phenotypes as an accessible diagnostic and prognostic marker to identify patients at risk.
Methods Neutrophils were isolated from peripheral blood and cells were analyzed under unstimulated and PMA-stimulated conditions following immunofluorescence staining for myeloperoxidase (MPO) and DNA. A total of 476 microscopy images were manually annotated to generate a training dataset for predicting changes in cell morphology during the progression of NETosis. Four different phenotypes were distinguished: cells with a (1) lobulated nucleus, (2) delobulated nucleus, (3) diffused NET and (4) spread NET; with 589, 839, 872 and 1618 cells annotated for the corresponding classes. Background images and various image transformations have been included to further increase the robustness of the model. A neural network for object detection, based on the YOLOv11 architecture, was trained on this dataset, and was applied to automatically identify neutrophils, allowing standardized quantification of NETosis.
Results Using this neural network, we demonstrated that PMA stimulation for 3 hours not only increased the number of NET-forming neutrophils but also significantly affected the preceding stages of NETosis—remarkably even more so than the overall increase in spread NETs (Fold increase in delobulated: 2.4 [±1.37] vs. diffused 14.74 [15.75] vs. spread 1.91 [±0.75] NETs; p=0.017). When analyzing neutrophils from patients with cardiomyopathy, we found that the relative increase in diffused NETs exceeded that of spread NETs, indicating an accumulation in intermediate stages of NETosis (33 [±36.2] vs. 10.94 [16.2]; p=0.023).
Conclusion We developed a Deep learning–based image analysis that allows objective quantification of neutrophil activation and identification of prognostically relevant phenotypes, with future potential for rapid point-of-care application in unprocessed blood samples. Our results show disproportionate increase in diffused NETs compared to spread NETs suggesting that not all neutrophils
proceed to full NET extrusion.
Figure: Representative immunofluorescence images illustrating the stages of neutrophil activation and NETosis. Shown are (a) a resting neutrophil with a lobulated nucleus, (b) nuclear decondensation, (c) diffuse NET formation, and (d) spread NET formation. Cells were stained for myeloperoxidase (MPO, red) and extracellular DNA (Sytox Orange, green).
