Classification of Human White Blood Cells Using Machine Learning for Stain‐Free Imaging Flow Cytometry

Abstract

Imaging flow cytometry (IFC) produces up to 12 spectrally distinct, information‐rich images of single cells at a throughput of 5,000 cells per second. Yet often, cell populations are still studied using manual gating, a technique that has several drawbacks, hence it would be advantageous to replace manual gating with an automated process. Ideally, this automated process would be based on stain‐free measurements, as the currently used staining techniques are expensive and potentially confounding. These stain‐free measurements originate from the brightfield and darkfield image channels, which capture transmitted and scattered light, respectively. To realize this automated, stain‐free approach, advanced machine learning (ML) methods are required. Previous works have successfully tested this approach on cell cycle phase classification with both a classical ML approach based on manually engineered features, and a deep learning (DL) approach. In this work, we compare both approaches extensively on the problem of white blood cell classification. Four human whole blood samples were assayed on an ImageStream‐X MK II imaging flow cytometer. Two samples were stained for the identification of eight white blood cell types, while two other sample sets were stained for the identification of resting and active eosinophils. For both data sets, four ML classifiers were evaluated on stain‐free imagery with stratified 5‐fold cross‐validation. On the white blood cell data set, the best obtained results were 0.778 and 0.703 balanced accuracy for classical ML and DL, respectively. On the eosinophil data set, this was 0.871 and 0.856 balanced accuracy. We conclude that classifying cell types based on only stain‐free images is possible with all four classifiers. Noteworthy, we also find that the DL approaches tested in this work do not outperform the approaches based on manually engineered features. © 2019 International Society for Advancement of Cytometry

Raman image-activated cell sorting

“…Finally, fluorescent labeling can lead to immunogenicity and introduction of xenobiotic compounds, such that human cells including human induced pluripotent stem cells (hiPSCs) 12 and chimeric antigen receptor T (CAR-T) cells 13 cannot be fluorescently labeled before in vivo use as cell therapies 14 . To overcome these limitations of fluorescent labeling, in silico labeling based on machine learning of numerous unlabeled (e.g., bright-field, phase-contrast) images has recently been shown as an alternative approach to identifying cellular state, interaction, and drug susceptibility 1 , 2 , 15 – 18 , but its application range is constrained by the lack of molecular specificity. We anticipate that the ability to sort cells based on unlabeled yet molecularly specific images with high throughput will significantly extend the utility and applicability of image-activated cell sorting and is, hence, expected to further advance single-cell biology and applications.…”

Deep-Learning Based Label-Free Classification of Activated and Inactivated Neutrophils for Rapid Immune State Monitoring

“…Similar work [ 24 ] has used the similar Resnet-18 model to classify label-free WBCs (neutrophils and monocytes) with the highest accuracy of 64.9%, but the accuracy of our method reached about 90%, as shown in Table 3 . The article pointed out that the reason for the unsatisfactory model was the imbalance of the dataset and low resolution of the obtained images.…”

Learning Deep Features for Stain-free Live-dead Human Breast Cancer Cell Classification

“…Additionally, data were augmented by random 90-degree rotations and vertical/horizontal flipping of each image. This type of data augmentation leads to better generalization performance 11 . We empirically found that fine-tuning network weights pre trained on Image-NET 25 performed significantly better than training from a random initialization.…”

“…In line with previous work 10,11 , we took advantage of well-known visualization techniques in order to gain further insight into the classifiers' automatically learned space to uncover their biological meaning. In particular, we applied a nonlinear dimensionality reduction technique suited for embedding high-dimensional data into a low-dimensional space, namely t-SNE 18 , which preserves local structures of the high-dimensional input space.…”