Fig. 1. CellSighter—a convolutional neural network for cell classification.

A Standard pipelines for cell classification take in multiplexed images and cell segmentation masks and generate an expression matrix. Cells in the matrix are annotated by rounds of clustering, gating, visual inspection and manual correction. CellSighter works directly on the images. B Imaging artifacts and biological factors contribute to making cell classification from images challenging. Segmentation errors, noise, tightly packed cells and cellular projections are easily visible in images, but hard to discern in the expression matrix. Scale bar = 5 µm. C CellSighter is an ensemble of convolutional neural networks (CNNs) to perform supervised classification of cells. D Comparison between labels generated by experts (x-axis) and labels generated by CellSighter (y-axis) shows good agreement. E Expert inspection of additional cells differing in classification between CellSighter and expert annotation. Inspection was performed blindly, without knowledge of the source of the label. F Comparing the recall of CellSighter (blue), XGBoost trained on the same labels as CellSighter (orange) and Clustering and gating (green). G For one field of view (FOV), shown are the protein expression levels (left), expert-generated labels (middle) and CellSighter labels (right). Source data are provided as a Source Data file.