Deep Learning of Morphologic Correlations To Accurately Classify CD4+ and CD8+ T Cells by Diffraction Imaging Flow Cytometry.

ABSTRACT

The two major subtypes of human T cells, CD4+ and CD8+, play important roles in adaptive immune response by their diverse functions. To understand the structure-function relation at the single cell level, we isolated 2483 CD4+ and 2450 CD8+ T cells from fresh human splenocytes by immunofluorescent sorting and investigated their morphologic relations to the surface CD markers by acquisition and analysis of cross-polarized diffraction image (p-DI) pairs. A deep neural network of DINet-R has been built to extract 2560 features across multiple pixel scales of a p-DI pair per imaged cell. We have developed a novel algorithm to form a matrix of Pearson correlation coefficients by these features for selection of a support cell set with strong morphologic correlation in each subtype. The p-DI pairs of support cells exhibit significant pattern differences between the two subtypes defined by CD markers. To explore the relation between p-DI features and CD markers, we divided each subtype into two groups of A and B using the two support cell sets. The A groups comprise 90.2% of the imaged T cells and classification of them by DINet-R yields an accuracy of 97.3 ± 0.40% between the two subtypes. Analysis of depolarization ratios further reveals the significant differences in molecular polarizability between the two subtypes. These results prove the existence of a strong structure-function relation for the two major T cell subtypes and demonstrate the potential of diffraction imaging flow cytometry for accurate and label-free classification of T cell subtypes.