Microfluidic systems are used to handle small amounts of liquid samples and are widely used in particle manipulation, including but not limited to particle detection and particle trapping [12,13]. Amiri et al [14] and Cha et al [15] proposed using microfluidic chips with a curved design for particle separation. Hasan et al [16] developed a polydimethylsiloxane microfluidic system to capture cancer cells. Solis-Tinoco et al [17] constructed a flexible microfluidic device to investigate live cell adhesion. Wu et al [18] used acoustic, electrophoretic, and hydrodynamic forces to prototype a microfluidic platform for cell separation. Beyond cellular manipulation, microfluidic platforms have demonstrated utility in the detection of H1N1 [19], H7N9 [20], and SARS-CoV-2 [21] viruses.

Combined with smartphones, microfluidics can enhance point-of-care diagnostic technologies. Chung et al [22] proposed a smartphone-based fluorescence microfluidic platform to detect norovirus in water samples, and Somvanshi et al [23] introduced a microfluidic paper-based aptasensor device, coupled with smartphone algorithms, to enable multiplexed detection of pathogenic bacteria. Most smartphone-based microfluidic platforms are limited to simple carriers containing biosamples due to the inaccessibility of droplet generators and syringe pumps [10]. Moreover, the use of 3D-printed microfluidics in SmartFlow offers advantages over traditional polydimethylsiloxane-based systems, such as greater flexibility, lower fabrication costs, and easier accessibility, as demonstrated by Au et al [24]. In comparison with existing smartphone-integrated systems, SmartFlow proposes using a 3D-printed, pump-free design that relies solely on gravity for fluid manipulation to eliminate the need for external components.

Our paper explores the methods to control microfluidic flow velocity for computer vision–based cellular analysis. Traditional flow speed control is achieved by syringe pumps, which are expensive and not ubiquitous in labs [25,26]. Therefore, previous work explored different natural power sources to drive flow in microfluidic systems [27]. Khor et al [28] and Xing et al [29] used liquid surface tension to control the flow speed. However, surface tension–driven techniques are hard to control and limited to liquid viscosity.

Prior work also focused on gravity-driven microfluidic systems. Goral et al [30] and Marimuthu and Kim [31] developed a communicating vessel–based gravity-driven system for cell cultures. Goral et al [30] used cellulose membranes to filter out unnecessary cells. Reis et al [32] used a glass siphon and multistep bioassays to quantify quantitative immunoassays. Kao et al [33] used communicating vessels to generate droplets for microfluidic analysis. Shin et al [34] designed a pressure-driven microfluidic system for colorimetric bioassays, and Gao et al [35] designed an external overflow unit to control the flow speed in a communicating vessel setting. Wang et al [36] achieved consistent flow speed using 2 siphons to keep the liquid surfaces. Limjanthong et al [37] tilted a table to power the flow in a microfluidic system for cultures of human-induced pluripotent stem cells.

Unlike gravity-driven cell culture systems and bioassays, computer vision–based microfluidic systems need to be observed under microscopes. Therefore, computer vision–based systems have stricter requirements for the flow velocity to ensure video quality, as we discussed previously. We first demonstrated that a high cell velocity could introduce motion blurriness to videos and lead to poor video quality. We then explored the feasibility of controlling the flow speed by using gravity to drive the flow and introducing a bottleneck design in the microfluidic channel. This technique can remove the high-speed cameras or syringe pumps used in previous microfluidic systems. We believe this is an important step to make smartphone-based microfluidic platforms more accessible.

Prior work applied computer vision algorithms to cellular analysis under microscopes. Red blood cell counting experiments can be automatically conducted using Hough transformation [38], and Lu et al [39] used cell signals in fluorescence assays to count cells in microfluidic droplets. Zeng et al [40] conducted a stained somatic cell counting experiment with traditional computer vision on dairy samples. Dima et al [41] and Shen et al [42] segmented florescence identities by setting a threshold on microscopy images.

As the prevalence of deep learning has surged in recent years, numerous studies have deployed deep learning models to analyze microfluidic cells. You Only Look Once (YOLO) [43] is a deep learning architecture proposed by Redmon et al [43] for object detection. Li et al [44] used YOLO on light scattering images to classify live and dead colonic adenocarcinoma cells. YOLO was also applied by Gardner et al [45] on inflorescence images to count cells in droplets. Moreover, Arjun et al [46] proved YOLO can be used to detect the mixing status of microfluidic droplets. High-speed cameras and convolutional neural networks can also be used in cell segmentation [47]. Lee et al [48] developed a user-friendly, fast deep learning model on cell sorting tasks in microfluidic systems. Furthermore, a convolutional neural network can be integrated into microscopic cell counting regression [49,50] and red blood cell morphology systems [51,52].

We applied dense optical flow algorithms in our study, given the ability of cells to flow through videos captured on our custom pump-free microfluidic platform. Compared with deep learning models, a dense optical flow algorithm had better interpretability, and it did not require a labor-intensive labelling process for cell segmentation. Additionally, we validated the effectiveness of our algorithms by manually counting the cells and estimating cell concentrations across samples of varying concentrations, which is considered the clinical standard for body fluid analysis.

We designed and fabricated 2 versions of microfluidic chips for this experiment, as illustrated in Figure 2 and Figure 3. Flow directions are highlighted with red arrows in the figures. Figure 2 shows the design of the straight version (version 1) of the microfluidic chip, where inlet and outlet were connected by a straight channel. The width of the channel was 2 mm, and the depth of the channel was 25 µm. However, we observed that the flow speed was highly sensitive to differences in the sample surface height, making it still challenging to control the flow velocity and maintain the video quality by adjusting the height difference.

To address this issue, we further designed a bottleneck version (version 2) of the microfluidic chip, shown in Figure 3. Introducing a bottleneck serves to slow down the flow velocity in 2 ways: First, according to Poiseuille's law, a narrower and longer channel introduces more resistance to the flow; second, the flow velocity decreases inversely with the cross-sectional area when the flow volume remains constant.

In this design, the channel remained at a depth of 25 µm. The central portion of the channel retained the same 2-mm width as the straight channel design to maintain comparability between the 2 versions. The bottleneck portion, however, was narrowed to 1 mm, reducing the cross-sectional area by one-half. This dimension was chosen to introduce sufficient resistance while still allowing cells to flow through the observation area without excessive slowing or clogging. Both microfluidic chips use slide covers sealed with ultraviolet glue to protect both the channel area and the microscope objective lens from contamination.

Figure 4 illustrates the SmartFlow setup for the cell flow velocity experiment. The system was modeled on the abstraction of communicating vessels, in which gravity serves as the power source. In communicating vessels, the liquid flows to reach uniform surface heights when a height difference is created.

As shown in Figure 4, 2 syringes served as reservoirs filled with the liquid sample, and the sample flows through the microfluidic chip. In this experiment, we used 2 versions of microfluidic chips, as illustrated in Figure 2 and Figure 3. A movable syringe adapter was designed to secure the syringe on one side, and it can be moved up and down manually as illustrated by the red arrow in the figures. The purpose for this was to generate the liquid surface height difference in the 2 syringes. A 21-gauge needle was used for the syringes, with an inner diameter of 0.514 mm and an outer diameter of 0.819 mm. Needles and microfluidic chips were connected with a plastic pipe with an inner diameter of 1 mm. The needle was used on the syringe because the needle was fitted for the plastic pipe so that it could enhance airtightness at the connection junction and effectively avoid sample leakage.

The smartphone was positioned directly in front of the eyepiece of the optical microscope using a commercial off-the-shelf (COTS) adapter. The COTS adapter allowed the smartphone to move in 3 axes. We tuned the position of the smartphone to ensure the view of the microscope from the eyepiece could be captured by the smartphone. Furthermore, the microfluidic chip was put on the stage of the microscope, and we adjusted the position of the microfluidic chip using the microscope’s stage controls and focus knobs to ensure the channel on the microfluidic chip was in focus. To initiate the flow, positive pressure was applied to one side of the syringe.

The objective of this experiment was to answer whether gravity can effectively drive the flow within the microfluidic chips, eliminating the need for external pumps or pressure systems (RQ1). In addition, we aimed to determine if the bottleneck design in the microfluidic chip reduces flow speed more effectively than straight channel designs, resulting in improved video quality for cell analysis (RQ2).

Before data collection, we captured images with gridded microscope slides (R1L3S3P, Thorlabs Inc) to establish a mapping correlation between image pixels and the actual scale observed through the microscope. We configured the microscope with an overall magnification of 1000 times, comprising a 10× magnification for the eyepiece and 100× magnification for the objective lens. The field number of the eyepieces was 22. We further calculated that the mapping correlation between the camera-captured image and the microscopy image was 5.37 pixels per µm.

We set up the entire system as shown in Figure 4 and diluted the sheep blood 10 times by mixing 0.5 mL of blood and 4.5 mL of saline. Data collection was performed using the application on the 2 chip versions, each with 7 different initial liquid surface height differences. For the straight chip, the height differences ranged from 0 cm to 0.7 cm, with increments of 0.1 cm in each trial. The height differences for the chip with the bottleneck design varied from 0 cm to 3 cm, with increments of 0.5 cm in each trial. We initially set the liquid surface to the same level, observed the flow speed at 0, and raised 1 side of the communicating vessels to the expected height difference. The cell flow velocity was estimated using a dense optical flow algorithm [53], in which dense optical flow could segment the moving objects and calculate the moving velocity and orientation of each pixel through adjacent frames in the videos.

The microfluidic chip design for this experiment was based on the abstraction in the flow velocity experiment setup. Figure 5 illustrates the cross-sectional and top views of the designed microfluidic chip used in this experiment. There are 4 chambers from left to right: the waste-sample reservoir, observation area, sample reservoir, and inlet. A slide cover was used to seal the observation area to avoid contaminating the objectives. In addition, a piece of plastic film was used to seal the waste-sample chamber, which could also be used to initiate the sample flow to the observation area. When using the microfluidic chip, the sample was poured into the inlet chamber, then the sample could automatically fill the sample reservoir. The plastic film was pressed to clear the air out through the hole (leftmost hole in Figure 5), then the hole would be blocked before the plastic film was released to initiate the flow.

Flow velocity was crucial to ensure the video quality, as fast flow velocity could lead to blurriness in the videos captured by the smartphone. Therefore, we designed an overflow mechanism in the sample reservoir chamber, so the maximum liquid surface height difference between the sample reservoir and the entrance hole to the waste-sample reservoir was 2.5 mm. In addition, we learned from the flow velocity experiment that a bottleneck design could slow down the flow to ensure the cell video quality, so the same principle was applied to the observation area. The thinnest width of the channel was 0.4 mm, and the diameter of the circle in the observation area was 2 mm.

In the cell concentration analysis experiment, a 3D-printed microfluidic chip was placed under the microscope, while the smartphone was placed in front of an eyepiece. This setup was much simpler than the previous setup, as the syringes and plastic pipes were discarded. This was achieved by abstracting the system setup for the flow velocity experiment into the design of the microfluidic chip for the cell concentration analysis experiment. During the experiment, cell samples with different concentrations flew through the microfluidic chip, and the videos were recorded using the Android app.

To conduct this experiment, we used sheep blood and diluted the blood with saline into 13 different concentrations. The red blood cell count in sheep whole blood is 9-15×10⁶ per microliter [54]. We focused exclusively on red blood cells, as the concentration of white blood cells is approximately 1000 times lower [54], making them barely observable after dilution. The concentrations used for the experiment are shown in Table 1, and the visualization is illustrated in Figure 6. The blood concentration is defined as the volume ratio between the blood and the sample. For example, a solution with a concentration of 0.1 has 1 microliter of blood contained in 10 microliters of the solution. In this way, the blood concentration before dilution is 1. In addition, the red blood cell count is defined by the number of red blood cells per microliter of the samples. These concentrations spanned a range that could be found in various body fluid tests, ranging from diluted blood samples in typical blood tests [55] to samples used for hematuria [56]. Videos were recorded via the mobile app through the eyepiece while the samples flowed through the microfluidic chip.

We used 2 ground truths to explore SmartFlow’s capability to estimate cell count and concentrations in the samples (RQ3). The first ground truth involved counting the cells manually from recorded videos. The second ground truth relied on the known blood concentrations for each sample during sample preparation.

Flow cytometers are used for cellular analysis in laboratory settings due to their high accuracy and precision. However, the high cost of these devices—ranging from US $100,000 to US $500,000—limits their accessibility, particularly in low-resource regions. Furthermore, operating a flow cytometer requires trained personnel, which further restricts their use to specialized laboratory environments. In contrast, the SmartFlow system is designed to be a low-cost, accessible alternative. The material cost for the 3D-printed microfluidic chip is approximately US $15, making it significantly more affordable. Additionally, our system is built using a COTS smartphone adapter, smartphone, and microscope, all of which are more commonly available in clinical and point-of-care settings.

The use of gravity-driven flow in our design offers practical advantages, particularly in low-resource or point-of-care settings. By eliminating the need for external pumps or complex fluid control systems, the design simplifies the overall setup and reduces the cost and technical complexity. This makes the system more accessible for use in clinical settings, where portability and ease of use are critical. However, the reliance on gravity also introduces some challenges, as the flow becomes more sensitive to height differences. In our experiments, we addressed this by introducing a bottleneck channel design to maintain a balance between flow speed and video quality. This design approach not only improves performance in our system but can also be applied to other microfluidic applications, such as cell cultures, to achieve better flow velocity control in similar gravity-driven setups.

The flow velocity was highly sensitive to changes in height, and due to the limitations of smartphone cameras, even small increases in height differences could lead to motion blur in the captured videos. The bottleneck design helped mitigate this sensitivity by controlling cell velocity. Since the flow rate (the volume of liquid passing through the channel per unit of time) remains constant throughout the channel, it is determined by the product of the cross-sectional area and flow velocity. In our design, the observation area has a larger cross-sectional area than the bottleneck, which has a smaller cross-sectional area. As a result, the flow velocity in the observation area is lower, and it becomes less sensitive to height variations. This reduction in sensitivity effectively minimizes motion blur, allowing for clearer, more stable video capture of the cells in the observation area.

We evaluated our system’s ability to predict blood concentrations using a 5-fold cross-validation method. The system achieved a mean absolute percentage error of 0.25 in predicting blood concentrations when compared with the ground truth. This demonstrates the system’s capability to perform concentration estimation. However, the accuracy of established technologies, such as flow cytometers, can serve as the ground truth in clinical settings. Therefore, we plan to explore methods to further improve the accuracy of our system by leveraging a more accurate flow volume and cell counting measurement.

Although SmartFlow can estimate cell concentrations, it currently only applies to single types of cells. Body fluids, such as urine and blood, contain multiple cell types, and changes in cell concentrations or morphology can be valuable indicators of various diseases. Therefore, we plan to conduct studies by further mixing different types of cells and integrating object detection and segmentation deep learning models into the system to support cell classification and segmentation.

The concentration ground truths we used were derived from sample preparation. Even if our system can estimate the flow velocity and the number of cells passing through the video, the cell velocity might differ from the liquid’s flow velocity. Moreover, the bubbles in the microfluidic channels could potentially impact the flow dynamics. Therefore, we cannot directly estimate cell concentration by dividing the number of observed cells by the amount of liquid that flowed through. However, clinical cell counting involves various criteria, including the number of cells per microliter and number of cells per high-power field after centrifugation under a microscope. We will conduct studies to compare our system against multiple clinical standards with more samples.

One assumption we made when calculating cell concentration was that the cells were evenly distributed in the fluid. However, over longer experiment durations, cells could settle down in the microfluidic chip, resulting in uneven cell distribution. This effect might influence the experiment results negatively when cell concentrations are low. Therefore, we plan to collect more data on low-concentration samples and evaluate the optimal experiment duration to obtain robust results.

Different biofluids can exhibit unique properties, such as viscosity, hydrophobicity, and hydrophilicity, which may interact differently with microfluidic material. Our system uses gravity to drive the flow and uses a bottleneck design to achieve better control over the flow velocity. These features are grounded in fluid dynamics principles, making them adaptable to fluids with diverse physical properties. By tuning the height differential and adjusting the bottleneck geometry, we believe the system can be optimized for a wide range of biofluids. Nonetheless, further testing is required to fully explore its applicability to fluids like saliva, urine, and cerebrospinal fluid, where more pronounced variations in viscosity or particulate matter may occur.