Deep Learning-Enabled Label-Free On-Chip Detection and Selective Extraction of Cell Aggregate-Laden Hydrogel Microcapsules

Abstract

Microfluidic encapsulation of cells/tissues in hydrogel microcapsules has attracted tremendous attention in the burgeoning field of cell-based medicine. However, when encapsulating rare cells and tissues (e.g., pancreatic islets and ovarian follicles), the majority of the resultant hydrogel microcapsules are empty and should be excluded from the sample. Furthermore, the cell-laden hydrogel microcapsules are usually suspended in an oil phase after microfluidic generation, while the microencapsulated cells require an aqueous phase for further culture/transplantation and long-term suspension in oil may compromise the cells/tissues. Thus, real-time on-chip selective extraction of cell-laden hydrogel microcapsules from oil into aqueous phase is crucial to the further use of the microencapsulated cells/tissues. Contemporary extraction methods either require labeling of cells for their identification along with an expensive detection system or have a low extraction purity (<≈30%). Here, a deep learning-enabled approach for label-free detection and selective extraction of cell-laden microcapsules with high efficiency of detection (≈100%) and extraction (≈97%), high purity of extraction (≈90%), and high cell viability (>95%) is reported. The utilization of deep learning to dynamically analyze images in real time for label-free detection and on-chip selective extraction of cell-laden hydrogel microcapsules is unique and may be valuable to advance the emerging cell-based medicine.

Keywords: cell microencapsulation; hydrogel; machine learning; microfluidic; transplantation.

Figure 1.

An overview of the deep learning-based detection and extraction system. (a) Flow chart of the overall system design and order of operations. (b) A 3D schematic illustration of the experimental setup. (c) A diagram of the microfluidic device showing its use for generating microcapsules and the deep learning-based label-free on-chip selective extraction of the cell aggregate-laden microcapsules. The microfluidic device consists of an oil channel inlet 1 (I1) for flowing an oil phase (containing CaCl₂), an I2 for flowing isotonic aqueous sodium alginate solution suspended with cell aggregates, and an I3 for flowing the isotonic aqueous extraction solution. Microcapsules are formed at the flow focusing junction (FFJ, i), gelled in the downstream gelling channel, and further flow into the detection region (ii) where images are taken for real-time detection. A deep learning-based detection program processes the images to determine if the microcapsules contain a cell aggregate or is empty. Once a cell aggregate-laden microcapsule is detected, the microcontroller is informed to turn the switch on, activating DEP force to extract the cell aggregate-laden microcapsule into the aqueous extraction channel (iii). Extracted microcapsules then flow down to the aqueous outlet (O1) where they are collected (iv). Non-extracted microcapsules continue to flow with oil to outlet O2. Scale bars: 100 µm.

Figure 2.

Characterization of the deep learning approach for label-free detection. (a) An illustration of the deep learning detection model including the backend architecture and a convolution (Conv) neural network using the single shot multibox detection (SSD), for detecting cell aggregate-laden microcapsules. (b) Quantitative data showing the image acquisition speed using the iPhone 7 camera. (c) Inference times of the deep learning detection models with three different backend structures.

Figure 3.

Characterization of the deep learning-enabled detection and selective extraction of cell aggregate-laden microcapsules. (a) Typical image sequence showing the selective extraction of a cell aggregate-laden microcapsule from oil into the aqueous extraction solution. The blue arrows indicate the flow direction, and the red/green box is the detection area. The green box indicates either no microcapsule or an empty microcapsule is present in the detection region, while the red box represents a cell aggregate-laden microcapsule is detected in the detection region. Scale bar: 200 µm. (b) Quantitative data of the deep learning-based detection efficiency and selective extraction efficiency of cell aggregate-laden microcapsules (n=3 independent runs with ~1000 microcapsules per run). (c) Quantitative data comparing the purity of microcapsules without and with the deep learning-based selective extraction (n=3 independent runs with ~1000 microcapsules per run for each condition). (d) Images of microcapsules collected from the device without and with selective extractions, showing the difference in purity between the samples. Scale bar: 500 µm. (e) Cell viability in both control cell aggregates without microencapsulation or extraction (control) and cell aggregates in microcapsules selectively extracted using the deep learning-enabled label-free method (n=3 independent runs with 7 cell aggregates per run for each condition).