Impedance flow cytometry empowered by ConvNet algorithm to differentiate bladder cancer cells based on electro-mechanical characteristics

As one of the top ten malignancies, bladder cancer (BC) is estimated to account for 6% of new cases and 4% of associated deaths in 2024,¹ making it a major global public health problem. Early detection of BC is effective in improving the 5-year survival rate of patients from 10% to 95%.² Urinary cytology-based liquid biopsy remains the gold standard for early diagnosis of BC with high specificity.^3,4 However, urinary cytology has low sensitivity and efficiency, because exfoliated BC cells can be covered up by normal epithelial cells and/or inflammatory cells.⁵ Therefore, it is urgently necessary to explore ways to rapidly differentiate between BC cells and epithelial cells and thus improve diagnostic efficiency and accuracy. In addition, distinguishing the grade of BC cells can provide valuable insights into cancer progression and thereby lead to improved treatment strategies.⁶ 

Single-cell analysis, as an effective means to study individual cells, plays a potentially significant role in elucidating cancer biology and has the promise of unveiling the heterogeneity of cancer cellular populations,⁷ thus providing possibilities for early diagnosis of cancer. Through the application of genomic, transcriptomic, and proteomics technology, single-cell analysis can identify cancer cells and even allow their molecular subtyping. These approaches, however, are complex, time-consuming, and expensive, hindering their widespread clinical use.^8,9

Recently, microfluidics has emerged as a powerful and cost-effective tool for single-cell analysis.^10–12 Microfluidic channels and single cells are comparable in size, and this facilitates the use of various biochemical or biophysical properties, including fluorescence,^13–15 Raman scattering,^16,17 and electrochemistry,¹⁸ for the detection and phenotyping of rare cells. In particular, the electrical and mechanical properties of cells are important label-free biophysical characteristics for comprehensive evaluation of cellular or subcellular physiological processes.^19,20 By integrating microelectrode-based sensing, microfluidic impedance flow cytometry (IFC) enables high-throughput measurement of electrical characteristics at single-cell resolution without the need for labeling²¹ and has therefore become widely used in cell phenotyping.^22–24 In addition, with appropriate design of the constriction channel, an IFC chip can collect details of the mechanical characteristics of individual cells.²⁵ Thus, constriction channel-based IFC can provide an efficient single-cell analytical technique for the simultaneous sensing of intrinsic cellular electrical and mechanical biomarkers. Along with advances in the physical construction of microfluidic systems, techniques for the analysis of IFC data have also improved. Artificial intelligence (AI)-based microfluidics possesses powerful data analysis and feature recognition capabilities, greatly enhancing differentiation between single cells.^26,27 For IFC, various algorithms (including those based on machine learning, deep learning, and neural networks) have been developed to perform single-particle analysis, differentiation between cells, and antibiotic susceptibility testing, bringing about great advances in the application of IFC in biosensing.^28–31 

Owing to the insignificant electrical and mechanical differences between various grades of BC cell populations and the large amount of impedance data generated by IFC, there is an urgent need for an automated and efficient data analysis method to improve the accuracy of differentiating between cells. However, AI-based IFC techniques still require further development to allow differentiation between cells of different grades. Furthermore, there is a dearth of systematic research on distinguishing between various BC cells using AI algorithm-assisted IFC detection. Here, we employ an IFC chip with contraction channel to sense the electrical and mechanical properties of BC cell lines. The differences between various bladder cells are investigated with the assistance of a convolutional neural network (ConvNet)-based deep learning algorithm. Both membrane and nuclear opacities and passage time parameters are defined to distinguish BC cells from epithelial cells by applying multiple frequencies. Additionally, the shape recovery parameter associated with the cells’ active deformation behavior is extracted and analyzed to investigate the electro-mechanical properties of cells for cell phenotyping. Then, by applying the ConvNet algorithm, the system is tested to differentiate between BC and epithelial cells, as well as different grades of BC cells from both spiked and patient samples.

The microfluidic IFC device was fabricated using a polydimethylsiloxane (PDMS)–glass technique and consisted of a constriction microchannel and microelectrodes. The microchannel was manufactured in PDMS using standard soft lithography,³² with cross-sectional dimensions of 10 μm width, 16 μm height, and 600 μm length. The widths of the microelectrodes were designed to be 15 μm, and the spacing between the side electrode and the excitation electrode was 55 μm. The microelectrodes were obtained by patterning 50 nm Ti and 150 nm Au on a BF33 glass wafer by a standard lift-off process.³³ After 4 min of oxygen plasma treatment, the microchannel was aligned and bonded to the glass substrate with the microelectrodes to form the IFC chip. The chip was then heat-treated at 90 °C for 40 min to ensure reliability. Finally, it was mounted on a printed circuit board (PCB) and wire-bonded to allow electrical measurements using an impedance analyzer.

The suspended cells were continuously pumped into the IFC chip by a pressure-based pump (Fluigent MFCS EZ, France) under a constant pressure of 20 mbar. A differential detection strategy was adopted using two pairs of microelectrodes, which were symmetrically placed near the entrance and exit of the constriction microchannel. An impedance analyzer (Zurich Instruments HF2IS, Switzerland) applied a sinusoidal alternating current (1 V) consisting of multifrequency components to the middle two electrodes. The differential current of the electrodes on both sides was pre-amplified 10 000 times by a pre-current amplifier (Zurich Instruments HF2TA, Switzerland) and then measured by the impedance analyzer with a sampling rate of 14.4 kHz.

The acquired electrical current data were filtered and detrended by homemade MATLAB program. Then, the signals of cells were identified and all the features were extracted as the datasets for convolutional neural networks (ConvNet) deep learning algorithm analysis, which was based on the TensorFlow 2.0 framework and programmed with Python 3.7. The designed ConvNet model included an input layer, three convolution layers, a full connection layer and an output layer. The convolution kernel was designed with a size of. The convolution depth was set to 64 layers and a rectified linear unit (ReLU) was used as the activation function. The dropout parameter in all of convolution layers was set 0.2 to optimize the number of unnecessary neurons and prevent overfitting..

The immortalized uroepithelial cell line SV-HUC-1 and BC cell lines (T24, EJ, and 5637) were acquired from the Chinese Academy of Sciences Committee Type Culture Collection Cell Bank (Shanghai, China). All the cancer cell lines were cultured in RPMI 1640 (Gibco BRL Life Technologies Inc., USA), and the SV-HUC-1 cells were cultured in F12K medium (Gibco BRL Life Technologies Inc., USA) containing 10% fetal bovine serum (FBS) (Gibco BRL Life Technologies Inc.) and 1% antibiotics (Gibco BRL Life Technologies Inc), which were incubated at 37 °C in a humidified incubator with 5% CO₂. After dissociation with trypsin, the suspended single cells were extracted and concentrated to 2 × 10⁵ cells/ml in 10 mM phosphate-buffered saline (PBS) by centrifugation at 600 rpm for the microfluidic IFC experiments.

Bladder tumor tissue and bladder epithelial tissue adjacent to tumor from the Second Hospital of Tianjin Medical University were contributed for this study. First, the tissues obtained by the surgeon were cut, grinded, and then digested by 2 mg/ml collagenase V (Solarbio, China) and 10 mg/ml dispase II (Solarbio, China) for 2 h to obtain single cells. The undigested tissue was then removed using a 70 μm filter membrane, and the dissociated cells were obtained by centrifugation at 450 g for 4 min. To prevent cell adhesion, 1 ml of 0.1 mg/ml DNase I (Solarbio, China) was added to the cell suspension for 10 min to remove free DNA. Finally, red blood cell lysate (Solarbio, China) was used to eliminate the influence of blood cells, and remaining dissociated cells were extracted and concentrated to 2 × 10⁵ cells/ml in 10 mM PBS for the microfluidic IFC experiments.

Images of cells passing through the contraction channel were taken by a high-speed camera (Photron UX50, Japan) mounted on an inverted fluorescence microscope (Olympus IX53, Japan) with a 20× objective lens (Olympus LUCPlanFLN20X, Japan).

All the experiments were conducted with at least three independent runs, and all data were reported as mean ± standard deviation.

Figure 1(a) shows a schematic of the microfluidic IFC system for differentiating BC cells and epithelial cells using a dual-parameter assay based on intrinsic cellular electrical and mechanical characteristics. As described in Sec. II A, the system incorporates two pairs of microelectrodes and a PDMS constriction channel. A differential sensing scheme is used to reduce impedance drift and ensure repeatability of measurements of the electrical signals of the cells as they flow between the microelectrodes within the microfluidic chip. According to the equivalent circuit model,³⁴ a single cell can be considered as equivalent to a capacitance and resistance in the dielectric field, and its impedance $Z̃cell$ is calculated by measuring the output signal of the side electrodes when a sinusoidal alternating current stimulus (1 V) is applied to the middle two electrodes. A cell’s electrophysiological behavior is closely linked to its biophysical properties and depends on morphological and subcellular characteristics. As shown in Fig. 1(b), in the electrical impedance analysis, three frequency components are employed to reveal the biophysical differences between cancer and epithelial cells. At low frequency (0.5 MHz), the membrane structure causes a cell to behave as an insulator, allowing estimation of its electrical diameter $Z0.5MHz1/3$⁠. When the excitation frequency is increased to middle range (5 MHz), the membrane capacitance is responsible for field dispersion. Changes in the internal properties of a cell (e.g., the structure of its nucleus or its cytoplasmic conductivity) are only manifested at high frequency (10 MHz).³⁵ In addition to the electrical impedance analysis, the system provides mechanical characterization of different cell types using the PDMS constriction microchannel (10 μm in width and 600 μm in length). As the cross section of the channel is smaller than the cell diameter, the application of a constant pressure (20 mbar) at the inlet will cause cells to deform on contact with the channel walls. The passage time of an individual cell passing through the channel is recorded by two pairs of electrodes [Fig. 1(c)], which reveals the cell deformability. Finally, by coupling electrical and mechanical signals to an AI algorithm for training and prediction, different cell types (including epithelial cells, low-grade BC cells, and high-grade BC cells) can be accurately differentiated, as shown in Fig. 1(d).

To test the feasibility of the proposed IFC chip for cell phenotyping on the basis of electrical and mechanical characteristics, over 8000 cells including three types of BC cells (from the low-grade 5637 cell line and the high-grade EJ and T24 cell lines) and one type of immortalized uroepithelial cells (SV-HUC-1 cell line) were studied and analyzed. Here, the electrical opacity, defined as the ratio of the impedance at a high frequency to that at a low frequency $ZHF/ZLF$ (including the cell membrane opacity $Z5MHz/Z0.5MHz$ and nuclear opacity $Z10MHz/Z0.5MHz$⁠), was used to take account of the effect of cell position and of variations in subcellular structure and content.^34,36 Figure 2(a) shows density heat maps of the membrane opacity $Z5MHz/Z0.5MHz$ vs the electrical diameter. It can be seen here that the opacity distribution of BC cells differs from that of epithelial cells, with the BC cells exhibiting a relatively small electrical opacity. This is to be expected, because cancer cells are known to have a rougher cell membrane with more folds, which will result in a larger membrane capacitance C_sm compared with the smoother membrane of epithelial cells.³⁷ It can also be seen that the membrane opacities of the high-grade BC cells (EJ and T24 cells) are smaller than that of low-grade BC cells (5637 cells), indicating that the membrane properties of BC cells are enhanced with increasing grade. When the signal frequency is increased to 10 MHz, at which the field short-circuits the cell membrane, polarization of the cell interior provides valuable information on cytoplasmic contents, such as the nuclear size.³⁸ As shown in Fig. 2(b), the difference in nuclear opacity between low-grade BC cells and epithelial cells becomes even greater, as does the difference between high-grade and low-grade BC cells. The lower nuclear opacity of BC cells can be attributed to their higher nucleus-to-cytoplasm (N/C) ratio and cell cycle turnover, which leads to increased cytoplasmic conductivity σ_cyt.³⁹ Furthermore, the N/C ratio of high-grade BC cells is larger than that of low-grade BC cells, leading to a lower nuclear opacity. The results shown in Figs. 2(a) and 2(b) also reveal that there is a difference between the membrane and nuclear properties of the two high-grade BC cell lines, with T24 cells having a greater membrane capacitance and smaller cytoplasmic conductivity. Intrinsic differences in mechanical metrics between the four cell lines were then analyzed on the basis of the measurements of nuclear opacity. The passage time measured by the two pairs of electrodes in the IFC system is known to be inversely proportional to cell deformability,⁴⁰ and the passage times of 32.20 ± 4.93 ms for SV-HUC-1 cells, 31.96 ± 5.42 ms for 5637 cells, 30.89 ± 1.81 ms for EJ cells, and 31.00 ± 2.78 ms for T24 cells [Fig. 2(c)] indicate that BC cells are more deformable than epithelial cells.

To further study the electro-mechanical properties of cells for cell phenotyping, the shape recovery parameters were extracted and analyzed. As shown in Fig. 3(a), when a deformed cell leaves the constriction channel, it gradually regains its shape owing to intracellular pressure generated by cortical actin.⁴¹ In comparison with nonmalignant cells, cancer cells recover their shapes more rapidly,⁴² returning to a morphology that is closer to their original one. This phenomenon affects the posterior peak signal in the bipolar peak of the IFC signal, with the cancer cells exhibiting a narrower peak width as well as a larger impedance, as shown by the red dotted box in Fig. 3(a). A high-speed camera captured the rapid recovery of BC cells as they left the channel, while the bladder epithelial cells underwent a much slower process [Fig. 3(b)]. We define two deformation recovery parameters W = w_recovered/w_native and P = Z_recovered/Z_native to facilitate study of the electro-mechanical properties of the four types of bladder cell lines. The results show that the distributions of the BC cell lines are significantly different from that of the epithelial cells [Fig. 3(c)]. The W parameter of the BC cells is much smaller than that of the epithelial cells, suggesting that the BC cells exhibit more active deformation behavior.

The impedance-related P parameter at different excitation frequencies was analyzed with the aim of exploring the differences in electrical response during active deformation recovery in different bladder cell lines. As can be seen from the density heat maps of electrical diameter vs P in Fig. 4(a), in the low- and mid-frequency bands (0.5 and 5 MHz), the density distributions of the BC cell lines are clearly different from that of the epithelial cell line. However, there is little difference between the four types of cell in the high-frequency band (10 MHz). The system mainly evaluates cell size and cell membrane information at low and mid-frequencies. Owing to the more rapid active deformation recovery of the cancer cells, their impedance is closer to that before deformation, resulting in their P parameter being much larger than that of the epithelial cells [Figs. 4(b) and 4(c)]. However, at the high test frequency of 10 MHz, the cell membrane is punctured, and the nucleus is sensed. Since the nucleus is not subjected to squeezing by the 10 μm constriction channel, there is no deformation recovery process for the nucleus. Therefore, there is no significant difference between the P parameters of the BC cells and the epithelial cells at high frequency [Fig. 4(d)]. Thus, in summary, W- and P-based analysis of the active deformation behavior of bladder cells is key to identifying new biomechanical markers for differentiation between different cell types.

The analysis of the IFC data allows for differentiation between BC cells and epithelial cells on the basis of their electrical and mechanical properties, but there is still some overlap of the distributions for the different cell types. To achieve precise differentiation between BC cells and epithelial cells, as well as between different grades of cancer cells, the ConvNet algorithm, as a representative deep learning algorithm, was applied to process IFC data from bladder cell lines. As shown in Fig. 5(a), all IFC data were first preprocessed and then inserted into a ConvNet algorithm model for learning and training. During the data preprocessing, 42 features at different excitation frequencies were extracted and reshaped into three-dimensional matrix data to match the algorithm, and then padded as a larger matrix to prevent loss of matrix edge information. The proposed ConvNet algorithm model was designed to contain an input layer, three convolution layers, a full connection layer, and an output layer (details of the model can be found in our previous work⁴³). During the model training process, all the IFC data were randomly divided into two parts: 70% as the training datasets and the remaining 30% as the cross-validation datasets to verify the reliability of the algorithm.

Figure 5(b) displays the confusion matrix of differentiation between cells based on the IFC data. The ConvNet algorithm accurately differentiates the three types of BC cells and epithelial cells with an accuracy rate of over 95%: a recognition accuracy of 95% for 5637 cells, 96% for EJ cells, and 97% for T24 cells. This demonstrates that the ConvNet algorithm performs well in distinguishing BC cells from epithelial cells, thus greatly improving the capability of the IFC technique for BC cell recognition. The IFC data from the three BC cell lines were further analyzed by the ConvNet algorithm, resulting in highly precise differentiation of BC cell subtypes. The accuracy rates for 5637, EJ, and T24 cells were 93%, 93%, and 92%, respectively. To further validate the reliability and practicality of the ConvNet algorithm, IFC data from spiked mixing samples of bladder cells were collected and input into the algorithm for prediction. First, spiked cell suspensions of 5637 and SV-HUC-1 cells mixed in 1:1 and 1:10 quantity ratios were tested using the IFC chip and predictions were made by the well-trained ConvNet algorithm. The prediction results are shown in Fig. 5(c), with recognition rates of 47.6% and 9.5%, respectively, for 5637 cells at the two different ratios. Similar studies were conducted on EJ and SV-HUC-1 cells, with recognition rates of 52.6% and 10.1% for EJ cells [Fig. 5(d)]. Spiked samples with three different quantity ratios of T24 and SV-HUC-1 cells (10:1, 1:1, and 1:10, respectively) were then examined, and the detection rates of T24 cells predicted by the ConvNet algorithm were 90.1%, 51.9%, and 11.7%, respectively. Finally, a spiked sample of three BC cell subtypes (from the 5637, EJ, and T24 cell lines) was prepared in a quantity ratio of 1:1:1 and tested on the IFC chip. The ConvNet algorithm predicted proportions of the three cell subtypes of 34.7%, 36.1%, and 29.2%, respectively. The cell recognition ratios of all the mixed samples described above are very close to the actual values, indicating the reliability of the ConvNet algorithm in assisting IFC data for differentiation between cells. Furthermore, the IFC technique assisted by the ConvNet deep learning algorithm has demonstrated its ability to differentiate between different grades of BC cells, and it thus has the potential for adoption as an effective tool for studying cancer cell heterogeneity.

The IFC technique combined with ConvNet algorithm-assisted data analysis has given satisfactory results in the analysis of cell lines. To further validate the capability of the proposed system for analyzing non-spiked cell samples, bladder tumor tissue composed of cancer cells and bladder epithelial tissue adjacent to the tumor were studied and the results were analyzed. First, as described in Sec. II D, the tissues obtained by the surgeon were cut, ground, and digested by collagenase and dispase II to obtain dissociated single cells, which were then transferred to PBS buffer for IFC detection after filtration, centrifugation, erythrocyte lysis, and removal of free DNA. The results of the IFC analysis shown in Figs. 6(a) and 6(b) reveal that the distribution of BC tissue cells is distinct from that of epithelial tissue cells. Compared with epithelial tissue cells, BC tissue cells have larger electrical diameter and smaller nuclear opacity. This indicates that the BC cells in tumor tissue are larger and have larger nuclei than nonmalignant cells. The ConvNet algorithm was further used for training and prediction of tissue cell data, and it gave differentiation accuracies of 96% and 94% for tumor tissue and epithelial tissue cells [Fig. 6(c)]. Thus, our proposed system for differentiating between cell types is able to accurately identify tumor cells in patient samples, thereby enabling label-free detection of BC cells.

In this study, a ConvNet deep learning algorithm-based IFC technique has been proposed to investigate the electrical and mechanical characteristics of bladder cells, including the SV-HUC-1, 5637, EJ and T24 cell lines. First, the parameters of membrane opacity and the nucleus opacity were defined and analyzed. Differences were found in the distribution of membrane opacity between BC cells and epithelial cells, indicating that the membrane capacitance of BC cells is greater than that of epithelial cells. Moreover, the significant differences in nuclear opacity distribution showed that the N/C ratio is largest in high-grade BC cells and lowest in epithelial cells, with an intermediate value in low-grade BC cells. The passage time, which is related to mechanical properties, was then investigated, and it was found that BC cells have stronger deformability. To further study the electro-mechanical properties of cells to enable cell phenotyping, the shape recovery parameters W and P of four bladder cell lines were extracted and analyzed. The obvious differences in these parameters indicate that BC cells have stronger active deformation behavior. In addition, was shown that with the help of a ConvNet algorithm, the IFC technique was able to effectively distinguish between BC cell lines and epithelial cell lines, with accuracy rates over 95%. Different grades of BC cells were also successfully distinguished, with accuracy rates of 93%, 93%, and 92% for 5637, EJ, and T24 cell lines, respectively. The test results for different proportions of spiked mixed samples were consistent with the actual values, indicating the reliability and effectiveness of the algorithm-assisted IFC data analysis in differentiation between bladder cell lines. Finally, it was demonstrated that the system could successfully distinguish bladder tumor tissue cells from epithelial tissue cells. All of these results prove that the ConvNet algorithm-assisted IFC technique has great potential for the early detection of BC in a noninvasive and label-free manner.

The authors gratefully acknowledge financial support from the National Natural Science Foundation of China (NSFC Grant No. 22076138) and the National Natural Science Foundation of China (NSFC Grant No. 62174119).

The authors have no conflicts to disclose.

Collection of patients’ tissue samples was approved by The Second Hospital of Tianjin Medical University Cancer Institute and Hospital Ethics Committee.

S. Z. and Z. Z. contributed equally to this work.

Shuaihua Zhang: Conceptualization; Methodology; Validation; Formal analysis; Data curation; Investigation; Writing – original draft; Visualization. Zhiwen Zheng: Conceptualization; Methodology; Investigation. Yongqi Chen: Software; Investigation. Zhihong Zhang: Supervision; Funding acquisition; Writing – review and editing. Ziyu Han: Conceptualization; Methodology; Validation; Formal analysis; Data curation; Investigation; Writing – review and editing; Visualization; Supervision. All authors have read and agreed to the published version of the manuscript.

The data that support the findings of this study are available from the corresponding authors upon reasonable request.

Shuaihua Zhang received the B.S. degree from Northeastern University, Qinhuangdao, China, in 2019. He is currently working toward a Ph.D. degree at Tianjin University, Tianjin, China. His research interests include biosensors, microfluidics fluorescence, and impedance flow cytometry.

Zhiwen Zheng received the Ph.D. degree from Tianjin Medical University, Tianjin, China, in 2023. He is currently working at The Second Affiliated Hospital of Anhui Medical University, Hefei, China. His main research interests are in urology.

Yongqi Chen received a B.S. degree from Xidian University, China, in 2020. He is currently working toward a Ph.D. in instrument science and technology at the College of Precision Instrument and Optoelectronics Engineering, Tianjin University, China. His research interests include microfluidic impedance cytometry, bulk acoustic wave resonator, and acoustic streaming.

Zhihong Zhang is Chief Physician and Full Professor at The Second Affiliated Hospital of Tianjin Medical University, mainly engaged in the urology clinic and in research on urological tumors.

Ziyu Han received the Ph.D. degree from Tianjin University, China in 2022. Currently, he is a postdoctoral fellow at the College of Precision Instrument and Opto-Electronics Engineering, Tianjin University. His research concerns microfluidic impedance cytometry, nanofluidic devices, and biosensing microsystems and platforms.