Machine learning classifies subcellular Golgi morphology from imaging flow cytometry during inflammatory activation: an exploratory study
Inflammatory activation remodels the Golgi apparatus to meet increased secretory demand, yet whether these structural changes carry computationally accessible information about immune cell state has not been explored. Here we show that machine learning applied to imaging flow cytometry images of the Golgi can distinguish lipopolysaccharide (LPS) treated from control cells. We trained models on Golgi images from human THP-1 monocytes before and 16 h after LPS stimulation in vitro, and from circulating myeloid cells of mice 24 h after intratracheal LPS or saline in vivo. In each dataset, the best-performing model was a deep learning model trained directly on images rather than a classifier using predefined morphometric features. Cross-domain performance varied with the model, transfer direction and use of labelled target data: several deep learning models trained on mouse images classified THP-1 cells above chance without changing model parameters. Grad-CAM highlighted regions within or adjacent to the GM130 fluorescence pattern. These findings indicate that GM130 images carry information associated with LPS treatment that supports classification within and across the two datasets. To our knowledge, this is the first application of supervised deep learning to imaging flow cytometry images of a subcellular organelle for inflammatory phenotyping, establishing organelle level image analysis as a potential complement to molecular assays for immune cell classification.