(M-493) Flexible Cell Sorting with Hydrogel Encapsulation Using Deep Learning

Introduction:: Current commercial cell sorters typically use sheath flow to align cells into a single profile and sort cells based on fluorescence signal or images. The single profile alignment limits the throughput and requires complex hardware and expensive equipment for high-speed sorting. The usage of high-speed sheath flow also generates high stress on cells. Some sticky cells such as monocytes or too many dead cells in the sample can interrupt or even clog the flow. Such cell sorter also usually requires a significant amount of starting cell number. Considering the yield, purity, and fluid dead volume, it is challenging to sort out cells of rare population such as subset of stem cells or circulating tumor cells in blood sample.

In recent years, there have been several image-based cell sorting (IBCS) techniques reported [1]–[3]. Yet, most image-based cell sorters face similar imaging challenges due to the high speed of the cells. Conventional cameras produce blurred images at a high cell flow rate up to 1 m/s in a typical microfluidic sorting device [4]. To capture blur-free images, complex approaches that scan the cells as 1D data with CCD line or laser imaging followed by subsequent image reconstruction have been used [5], [6], which are time-consuming and require precise synchronization.

To address these drawbacks, we propose a new microfluidic cell sorting device that can process a massive number of cells at the same time through image-based machine learning, segmentation, and selective encapsulation.


Materials and Methods:: As the first step of the proposed method, we demonstrate the feasibility of using selective encapsulation of target in sorting cells. A large microfluidic chamber accommodates the beads/cells mixed with hydrogel for imaging. The chamber contains a narrow inlet and outlet to ensure minimum sample to air interface where beads/cells can create congestion and cause inaccurate detection. A high-precision stage is used to move the microfluidic device and image the entire chamber. A color CMOS camera (Ximea CB120CG-CM) is used to capture high-resolution images of the sample and a YOLOv5 based deep learning object detection model is used to detect target and isolate it from the sample through selective encapsulation.

The strength and flexibility of the method originates from imaging. High-resolution RGB images captured in this method are blur free and contain a high level of information. As the sorting method does not rely on sophisticated control, any target can be encapsulated and isolated. So, targets can be determined based on size, morphology, color, or even properties correlating to nearby objects unlike to conventional cell sorters.

For this experiment, we first used two types of beads (smooth and rough) and trained them in YOLOv4 using 1000 images. Then using the trained model, we are able to separate rough beads from smooth beads. We then separate beads based on size. We later train another model with 1000 brightfield images of PBMC and PC3 cells and show high prediction accuracy as well.


Results, Conclusions, and Discussions:: The results are summarized in Figure 1(a). For both beads and cell training, the mean average precision (mAP) reaches a plateau at ~90 epochs. The confusion matrix shows 98% and 99.6% true positive (TP) and false negative (FN). In the case of differentiating between PBMC and PC3 cells it shows over 99% accuracy for TP and FN. The accuracy data has been measured from validation set which is not known to the model. The cell images are less blurry compared to other high-speed cell sorters as seen in Figure 1(b).

The size-based sorting of the beads shows how this method can be used to target any size of target with high accuracy. When the desired target size was set to ~17-18 µm, the output shows no bead outside the size range while losing a total of only 0.3% of the 3381 target beads. This can be compared to monodispersed microparticles.

The proposed method can be an approach in the cell sorting industry as deep learning is being increasingly implemented in similar applications. This method works with extremely low starting volume. For the experiment mentioned above, only 50 µL of the sample was dispensed in the inlet. As there is no sheath flow involved, it not only avoids high sheer-stress but also avoids downstream cell loss due to large dilution of the cells. Additionally, there should not be any size limitation of the target cells meaning, this can be used for sorting cell spheroids and organoids which is often incompatible with conventional sorters due to their size.


Acknowledgements (Optional): :

References (Optional): : [1] R. S. W. Thomas, P. D. Mitchell, R. O. C. Oreffo, H. Morgan, and N. G. Green, “Image-based sorting and negative dielectrophoresis for high purity cell and particle separation,” Electrophoresis, vol. 40, no. 20, pp. 2718–2727, Oct. 2019,

[2] M. Sesen and G. Whyte, “Image-Based Single Cell Sorting Automation in Droplet Microfluidics,” Sci. Rep., vol. 10, no. 1, pp. 1–14, May 2020,

[3] K. Gardner, M. M. Uddin, L. Tran, T. Pham, S. Vanapalli, and W. Li, “Deep learning detector for high precision monitoring of cell encapsulation statistics in microfluidic droplets,” Lab Chip, vol. 22, no. 21, pp. 4067–4080, Oct. 2022,

[4] C. A. LaBelle, A. Massaro, B. Cortés-Llanos, C. E. Sims, and N. L. Allbritton, “Image-Based Live Cell Sorting,” Trends Biotechnol., Nov. 2020,

[5] S. Ota et al., “Ghost cytometry,” Science (80)., vol. 360, no. 6394, pp. 1246–1251, Jun. 2018,

[6] A. Isozaki et al., “A practical guide to intelligent image-activated cell sorting,” Nat. Protoc., vol. 14, no. 8, pp. 2370–2415, Aug. 2019,