Fig. 3: Image annotation of mass data improves measurement fidelity.

a Schematic representation of the multi-step image classification approach implemented in the workflow depicted in Fig. 1. An input image is first classified as accepted or rejected with a binary CNN classifier, and subsequently, all accepted particles are classified as either an intact or permeable cell and all rejected particles are classified as either an aggregate or debris with two additional binary CNN classifiers. The inset images are representative of each class. The parenthetical percentages listed at each binary decision node indicate the cross-validated model performances, listing the precision and recall values, respectively (Methods). b Violin plots showing the mass distributions for a full population of cells without image classification (white) as well particles within the population that were classified as aggregates (purple), intact cells (gray), permeable cells (pink), or debris (blue) for a human lung cancer cell line model (PC9) and a human multiple myeloma cell line (MM1S). c Single-cell image classification reduces the noise in mass response signal by reducing the sampling error. As a demonstration, we randomly sample 1000 cells 100 times from a measured condition with and without image curation and calculate the mean mass response within these two sets representing the sampling error with and without image curation. When this is repeated for all the measurements presented in this paper, including 3222 measured conditions across 13 different instruments, we find that image curation reduces the sampling error by 16.2% on average. An example pair of mass distributions with (gray) and without (white) image curation is shown, indicating the presence of cell aggregate and debris populations in the original measurement set, which are subsequently removed by the image curation process.