Biomedical Microdevices

, Volume 14, Issue 1, pp 83–93

A novel silicon membrane-based biosensing platform using distributive sensing strategy and artificial neural networks for feature analysis

Authors

  • Zhangming Wu
    • School of Engineering and Applied ScienceAston University
  • Khujesta Choudhury
    • School of Engineering and Applied ScienceAston University
  • Helen R. Griffiths
    • School of Engineering and Applied ScienceAston University
  • Jinwu Xu
    • University of Science and Technology Beijing
    • School of Engineering and Applied ScienceAston University
Article

DOI: 10.1007/s10544-011-9587-6

Cite this article as:
Wu, Z., Choudhury, K., Griffiths, H.R. et al. Biomed Microdevices (2012) 14: 83. doi:10.1007/s10544-011-9587-6

Abstract

A novel biosensing system based on a micromachined rectangular silicon membrane is proposed and investigated in this paper. A distributive sensing scheme is designed to monitor the dynamics of the sensing structure. An artificial neural network is used to process the measured data and to identify cell presence and density. Without specifying any particular bio-application, the investigation is mainly concentrated on the performance testing of this kind of biosensor as a general biosensing platform. The biosensing experiments on the microfabricated membranes involve seeding different cell densities onto the sensing surface of membrane, and measuring the corresponding dynamics information of each tested silicon membrane in the form of a series of frequency response functions (FRFs). All of those experiments are carried out in cell culture medium to simulate a practical working environment. The EA.hy 926 endothelial cell lines are chosen in this paper for the bio-experiments. The EA.hy 926 endothelial cell lines represent a particular class of biological particles that have irregular shapes, non-uniform density and uncertain growth behaviour, which are difficult to monitor using the traditional biosensors. The final predicted results reveal that the methodology of a neural-network based algorithm to perform the feature identification of cells from distributive sensory measurement has great potential in biosensing applications.

Keywords

BiosensorsMicroscale membraneDistributive sensingNeural networkEndothelial cell line

1 Introduction

Research into the use of biosensors for the detection of various biological particles/molecules has received extensive interest in recent decades, due to the rapid progress of micro/nano technologies. A biosensor usually consists of a bioreceptor and a sensing transducer, in which the bioreceptor is the interface that the biosensor interacts with the biological environment and the transducer is used to convert the physical/ chemical information of the biological particles into a measurable signal (Mohanty and Kougianos 2006). Microcantilevers are the most widely used transducer in mechanical-type biosensors, due to their ultra-small size, high sensitivity and label-free biological application by surface functionalization (Raiteri et al. 2001). This kind of biosensor can be fabricated in arrays of microcantilevers (Carrascosa et al. 2006) and can be integrated into a CMOS-based microsystem. In general, microcantilever can work under two different modes: static mode and dynamic mode (Lavrik et al. 2004; Wee et al. 2005). In a static mode, surface stresses are accumulated after binding biological particles to the microcantilever surface and they will inevitably increase the static deformation of microcantilever. The dynamic mode uses a microbalance approach, which detects surface-attached mass using resonant frequency shift. However microcantilever based biosensors suffer from low sensitivity in liquid environment and fragility in practical operation (Xu et al. 2008).

Micromachined membranes (plate/diaphragm) have gradually become a promising mass sensing structure to replace the microcantilever in recent years (Nicu et al. 2005; Nicu and Ayela 2007; Ayela et al. 2007). Compared with microcantilevers, micro-membranes potentially have larger sensing area, higher sensitivity in liquid and less fragility. Moreover, it may have the same advantages as the microcantilever in the application of mass sensing. Some researchers have made attempts to apply micro-membranes for biological detection. For examples, Carlen et al. (2006) designed a micromachined surface stress sensor based on a thin suspended crystalline silicon circular plate to detect the bending behaviour caused by vapor phase chemisorption of the alkanethiol monolayers. Xu et al. (2008) developed a piezoelectric membrane-based biosensor array for immunoassay applications.

Distributive sensing techniques have been widely used to monitor and reconstruct the static deformation or dynamic responses of conventional structures, in the field of vibration control, damage detection and biomedical analysis etc. One famous instance is a beam-like or plate-like smart sensing surface with few distributive tactile sensors. The sensors are placed at selected locations and used to collect the data of surface deformation. Any change upon the sensing surface can result in a corresponding change of measurements in each sensor. The features or the properties of contacted object are related to the sensory data. Advanced nonlinear feature analysis methods, for example artificial neural networks, can be applied to infer the properties of the contacted subject. These kinds of tactile sensing surfaces have been successfully applied to determine a description of force loading (Ma and Brett 2002), localisation of a contacting subject (Cowie et al. 2006) or even human gait analysis (Elliott et al. 2007). Apparently the design of integrated microsystems for biosensing can also borrow the concept of the smart sensing surface, in which multi-dimensional signals rather than single output from the sensing surface can be collected and used for post-processing. Therefore it has good potential to extract more information such as distribution or pattern of biological targets, rather than just the mass variations.

Frequency response function (FRF) is one of the most useful ways to represent the dynamics of a rectangular membrane. Cell adhesion on a membrane surface results in the change of its mass, stiffness and damping, all of which can be reflected in a FRF. By comparing the FRFs of a sensing membrane with and without cell attachment, we can identify the features of adhesive cells. Many researchers have successfully employed neural networks on the measued FRF data for structural health monitoring and damage detection (Chaudhry and Ganino 1994; Zang and Imergun 2001; Ni et al. 2006; Lee and Kim 2007). This paper presents the successful cell monitoring application. Before applying FRF data into training a neural network, the size of FRF data has to be reduced. Many methods exist to perform the data dimension reduction, such as sub-dataset (Chaudhry and Ganino 1994), modal analysis (Levin and Lieven 1998) and principal component analysis (Zang and Imergun 2001). In this paper, Karhunen–Loeve decomposition is used for FRF data dimensionality reduction.

The purpose of this paper is to apply a micromachined rectangular membrane as biosensing platform using the distributive sensing method. This novel approach is first in the field of biosensing, which paves the way for developing more advanced biosensors with high accuracy and multiple functions. The paper is organised as follows. Section 2 presents the fabrication methods of biosensing platform including the distributive piezoresistive sensors and PZT actuators for self-actuation and self-sensing. Section 3 presents the process of biological testing. Experimental tests of biosensing micro-membranes are used for detection of EA.hy 926 endothelial cell lines in a natural liquid environment. EA.hy 926 is a well-established human endothelial-like immortalised cell line that exhibits adherence and migratory characteristics, resulting in non-uniform shapes in culture. Dynamic responses of the micro-membrane at a few specific points are measured and recorded. Such information forms a set of distributive sensory data. In analysing the sensory data, first a shift of resonance frequency at each measured mode is used to perform a preliminary estimation of the cell density. It is found that frequency based indices alone are unable to accurately reflect the attached cell distribution on the sensing surface. Finally, in Section 4, a Back-Propagation (BP) neural network with one hidden layer is trained to recognise the cell distribution from the distributive sensory data of a series of repeated bio-experiments. It shows precise prediction of cell density by using this neural network model.

2 Fabrication of membrane biosensing devices

The silicon membranes were fabricated using the standard micromachining techniques from silicon on insulator (SOI) wafers with 5 or 3 μm thickness active silicon layer. The membrane was created by inductively coupled plasma (ICP) using the Deep Reactive Ion etching (DRIE) process from the back side of SOI wafer, stopping at the buried oxide layer. Boundary conditions of the membrane were also defined by DRIE from the top side of the wafer, using the buried oxide as stop layer. The buried oxide layer was finally removed to form the boundary holes. Three different boundary conditions of the micro-membranes were fabricated and tested: two opposite edges clamped and the other two edges free (C-F-C-F), cantilever (C-F-F-F) and all edges clamped (C-C-C-C). All of the membranes are designed to be square and with lengths of 100, 200 or 300 μm.

Different types of integrated microsystems were also designed and manufactured, based on a square sensing membrane with distributive piezoresistive sensors and PZT actuators. Figure 1 demonstrates one of such microsystems, in which a 100 μm square silicon sensing membrane is used. Such a microsystem enables the device to be capable of self-sensing and self-excitation. This microsystem can be embedded into an electronic circuit to build a lab-on-chip system. The fabrication procedures of these microsystems are similar to Lu et al.’s work on microcantilevers (Lu et al. 2008, 2009).
https://static-content.springer.com/image/art%3A10.1007%2Fs10544-011-9587-6/MediaObjects/10544_2011_9587_Fig1_HTML.gif
Fig. 1

SEM image of an integrated microsystem using a 100 μm square membrane and attaching with distributive piezoresistive sensors and PZT actuators

For the fabrication of distributive piezoresistive sensors, a 500 nm-thick poly-silicon layer was deposited onto the oxidised device layer of a SOI wafer by low pressure chemical vapour deposition (PCVD). This layer was then doped by ion beam implantation using a 50 Kev Boron source giving a doping density of 1e15 to enhance the piezoresistive deflection sensitivity. The sensor shapes were formed by photo-lithography and subsequent reactive ion etching (RIE).

In the PZT film fabrication, a sandwiched structure of a 100 nm-thick Pt/Ti bottom electrode, a 1 μm PZT film and a 100 nm-thick Pt top electrode was deposited on the SOI. The top and bottom electrodes were deposited by evaporation using e-beam evaporator systems, the deposited PZT was deposited as a spin on sol-gel which is then annealed to produce the required PZT film. The top and bottom electrodes are patterned and etched by ion beam milling.

The redundant PZT material was wet etched using solution of BHF:HCl(38%):NH4Cl(98%):H2O = 1:2:4:4, with the top electrode and resist acting as the mask. Then an additional 500 nm-thick electrode (Aluminum) layer was deposited and patterned to form the electric pads and wires. All the redundant SiO2 and other materials were etched and removed from the top surface of SOI wafer, especially on the define domain of membrane. Finally, back side silicon and buried SiO2 layer were etched and removed to form the membrane structure.

The electronic part of this system (electrode wires, gold pads, and connecting probes) are sealed with biocompatible material (silicon). Each device is packaged into a 46 pin DIL (Dual in-line) package for experimental operation. The signal flow (input and output signals) is processed via the NI USB-5132 Oscilloscope and a self-made circuit board, in which the Wheatstone bridge and amplifier are used to enhance the response signals from piezoresistive sensors. As the circuit board contains matching resistors, the four piezoresistive sensors can work either individually in a quarter bridge or work all together in a full bridge. The PZT actuators are directly driven by 2 ∼ 5 V pseudo-random excitation signals (Wu et al. 2010), which is an ideal approach for the dynamic testing of micromembranes.

3 Biological experiments

3.1 The process of bio-experiments

The human hybrid EA.hy 926 cell used in this paper is derived from the fusion of the human umbilical vein endothelial cells with A549/8 human lung carcinoma cell line. EA.hy 926 is a stable human endothelial cell line that expresses highly differentiated functions characteristic of human vascular endothelium. Human EA.hy 926 endothelial cell lines are maintained in 30 ml Dulbecco’s Modified Eagle’s Medium (DMEM), supplemented with 10% FBS, streptomycin 100 μg/ml and penicillin 100 U/ml, and 10 ml HAT (100 μM hypoxanthine, 0.4 μM aminopterin, 16 μM thymidine). Cells were cultured in an incubator at 37°C with an atmosphere of 5% CO2 and 95% air. Cells were grown in a 75 cm2 flask and passaged when reaching ∼90% confluence. Once cells roughly reached 90% confluence the media was removed and the cells washed with 5 ml phosphate buffered saline (PBS). The process of passage of EA.hy 926 cells is that briefly cell culture media was removed from the cells and cells were then washed with 10 ml sterile PBS until the media appears without color. EA.hy 926 cells were then detached by the addition of 2.5 ml trypsin with a 3 min standard incubation. Cell clusters were dispersed for uniform distribution by repeated pipetting with 5 ml new DMEM media.

The Fig. 2(d) shows a LSM image (Laser Scan Microscopy, captured by ZEISS LSM 510 Meta Confocal Microscopy) of the endothelial cells coated on the surface of a micro-membrane. It can be seen that those endothelial cells were tightly adhered to the silicon surface showing a typical spreading pattern.
https://static-content.springer.com/image/art%3A10.1007%2Fs10544-011-9587-6/MediaObjects/10544_2011_9587_Fig2_HTML.gif
Fig. 2

Endothelial cells coating on the surface of a micro-membrane: (a) shows a silicon die (membrane) inside in a petri dish, (b) shows the same after a period of incubation, (c) is the dynamic testing device, (d) is the laser scanning image of endothelial cells coating on the surface of a micro-membrane

The seeding of biological experiment is separated into two phases: seeding a certain amount of cells on the membrane (Fig. 2(a) and (b)) and measurement of the corresponding dynamics of this membrane. The dynamic testing device is illustrated in Fig. 2(c). Identical micro-membranes were repeatedly used several times for obtaining a batch of experimental results with different densities of cells. Each experiment was performed according to the following work flow:
  1. 1.

    Initially, silicon micromembranes were cleaned and sterilised using washes (ethanol and acetone mixture), autoclaving and UV light irradiation.

     
  2. 2.

    Before seeding cells on the micro-membranes, the cell density of suspension during the process of passage was established. The numbers of viable cells were estimated by taking 20 μl of the cell suspension and mixing it with a 20 μl trypan blue. Cells count was then performed from this new mixture by using improved Neubauer haemocytometer. Once the cell density was established, a 5 ml cell suspension of EA.hy 926 cells of known density is made up using the media. By controlling the incubation time (1 ∼ 4 days), various cell density and distribution on the membrane surface can then be achieved.

     
  3. 3.

    Cell distribution on the membrane sensing surface is recorded using a LSM image. The density or distribution of cells can be quantitatively estimated based on this LSM image.

     
  4. 4.

    The dynamics of membranes with adherent cells are measured. The FRF data for each specific micro-membrane with cells and without cells are compared to infer the information of cells, which is recorded in the LSM scanned images.

     
  5. 5.

    Finally, the cells are removed from the surface of micro-membranes following the same procedure as the first step. The re-sterilised micro-membrane can be used for the next experiment.

     

3.2 Experimental results

Figure 3 illustrates the frequency response functions (FRFs) of three different types of micro-membranes under three different cell densities. The most dominant change of the dynamics of membrane induced by cell-loading is the shift of resonance frequencies. As the first mode shapes remain almost constant (Wu et al. 2009), and the amplitudes of each FRF were self-normalised with respect to the amplitude of first resonant mode, relative amplitudes of resonant modes are found to be significantly varied after the cell loading. It means that additional mass loading of attached cells on the surface of membrane also results in the distortion of vibration shapes. The mass or quantity of target cells can be estimated through the detection of the shift of resonance frequencies. Equation 1 demonstrates the relationship between mass change and frequency shift of a dynamic system, under the assumption that the stiffness remains constant. This approach has been widely used in the microcantilever based biosensors.
$$ f=\frac{1}{2\pi}\sqrt{\frac{k}{m}}, \;\;\;\; \frac{\Delta m}{m} = \frac{k}{4\pi^{2}}\left(\frac{1}{f^{2}_{1}}-\frac{1}{f^{2}}\right) \approx 2\frac{\Delta f}{f} $$
(1)
where f and f1 represents the resonance frequency of micro-membrane without cells and with cells, respectively; k and m are the stiffness and mass of micro-membrane, respectively; Δf is the frequency shift of micro-membrane due to cells adhesion; Δm is the mass change of micro-membrane, which is equal to cells mass.
https://static-content.springer.com/image/art%3A10.1007%2Fs10544-011-9587-6/MediaObjects/10544_2011_9587_Fig3_HTML.gif
Fig. 3

FRFs of three different types of micro-membranes under three different cell densities. Above LSM images of endothelial cells coating on the surface of square micro-membranes. Below normalised amplitude of above specified micro-membrane. (a) 100 μm square C-F-F-F micro-membrane, (b) 200 μm square C-F-C-F micro-membrane, (c) 300 μ m square C-C-C-C micro-membrane

Comparing the changes of FRFs presented in Fig. 3, it is concluded that different types (dimension and boundary conditions) of the rectangular silicon micro-membranes reflect very different biosensing performance. It implies that the first type membrane (a 100 μm square C-F-F-F) has highest sensitivity among the three membranes, in terms of resonance frequency shift. It is also noted that most of the measured dynamics (FRF data) of micro-membranes involving cell attachment have suffered with nonlinearity. In general, the experimental results shown in Fig. 3 demonstrate the great potential of micro-membranes in biosensing, even when they are immersed in a high-damping liquid environment (Alava et al. 2010).

3.3 Preliminary analysis

Two resonant frequency based indices (Eq. 2) are utilised to perform a preliminary analysis on the experimental results in this paper. FDRn (Frequency Difference Ratio) is evaluated as the normalised resonant frequency difference between the cell-loaded and cell-free membrane at each measured resonance mode. AFDR is the average of all measured FDRn.
$$ FDR_{n}=\frac{\Delta f_{n}}{f_{n}}, \;\;\;\; AFDR=\frac{1}{N}\sum^{N}_{n}FDR_{n}$$
(2)
The indices of FDRn and AFDR evaluation were performed on three batches of bio-experimental results using three different micro-membranes, which are all approximate 200 μm square C-F-C-F membranes. The three micro-membranes are labeled as No. I, No. II and No. III respectively. In each batch of the experiment, an identical membrane was repeatedly used four times and the cell culture density was gradually increased from to 25 × 103/μl to 200 × 103/μl. Figure 4 illustrates the trends of the FDRn with increasing the amount of cells on each tested micro-membrane. The right bottom image in Fig. 4 compares the AFDR index of these three micro-membranes in each batch of experiments.
https://static-content.springer.com/image/art%3A10.1007%2Fs10544-011-9587-6/MediaObjects/10544_2011_9587_Fig4_HTML.gif
Fig. 4

(a)–(c) Above is the FDR trends of the three micro-membrane (No. I–No. III) in four independent biological experiments; below show the LSM images of typical cell growth observed at the above four experiments. (d) Shows AFDR trends of the three micro-membranes in each batch of bio-experiments

First of all, some trends of the index FDRn at one or two modes are not coherent with the increase of cell quantity. This phenomenon is quite different with the bio-experimental results of a microcantilever, where the FDR0 of its fundamental mode always has a linear relationship with cell number (Ilic et al. 2004; Gupta et al. 2004). The potential reasons of this phenomenon are: (a) Micro-membranes usually have much larger sensing area and carry many more cells than a microcantilever in the bio-experiments. Apart from mass change, the accumulation of cells may also result in change of structural stiffness. In such cases, the linear relationship of FDR will be violated. (b) The bio-experiments presented in this work for micro-membranes are conducted in real cell culture media rather than an ideal vacuum environment. (c) Nonlinearity of the dynamics of submerged micro-membranes with randomly distributed cells exists in most experimental measurements.

On the other hand, index AFDR is capable of giving an approximate prediction of the amount of cells. The sensitivity of AFDR on these three micro-membranes is quite different. The values of AFDR for No. I and No. II membranes are very close, but that of No. III is much lower. This is due to the fact that No. I and No. II membranes were taken from the same wafer, while No. III is from the other one. It reveals that using the index AFDR for the micro-membrane as a biosensing platform is not a robust method. Calibration on such a biosensing device is probably required before the estimation on cell density.

Considering the submerged sensing membrane as a general oscillation structure, resonant frequency can be approximately determined only by its stiffness and mass, the first equation in Eq. 1. If one assumes the system stiffness is a constant, the mass change ratio is proportional with frequency change ratio as shown in second equation of Eq. 1. It is therefore believed that indices FDRn and AFDR are able to roughly reflect the cell density. However in realistic situations cell attachment would also affect the stiffness of sensing micro-membrane more or less, especially the endothelial cells. Therefore FDRn and AFDR are less accurate for indicating the cell density.

4 Neural network method

On the whole, resonant frequency based indices (either FDRn or AFDR) are only able to predict the cell density with very limited accuracy. It is mainly due to the complication and nonlinearities of micro-membrane sensing system. Other algorithms are desired to perform more accurate and reliable identification on cell distribution from the measured dynamics data. In this section, a simple attempt that using an artificial neural network technique to build the relationship between the sensory data and cell distribution is carried out.

4.1 Quantitation of cell density

In the above experimental results, LSM images were used to intuitively present the cell population in the micro-membrane sensing domain. However a quantitative index is also necessary to indicate the amount of cells for a more precise analysis. This is especially true for endothelial cells, the number of which is difficult to count. A simple image processing procedure was carried out on each LSM image to convert it into a binary image using the MATLAB Image Processing Toolbox. Initially the LSM image is loaded and a most clear layer is selected for the following processes, as the LSM image taken under the reflection mode usually contains three layers. Then the background image of this LSM image is created by the morphological opening technique. Afterwards the background image is subtracted from the original image and the image contrast is enhanced, in order to highlight the area occupied by cells. Finally the corresponding binary image is created, in which the background is black and the parts of implanted cells are white. Therefore the cells population on the sensing domain can be approximately evaluated by the white area ratio in this binary image. This ratio is called cell density ratio (CDR) in this paper. Figure 5 demonstrates the results of this evaluation processes on four different LSM images, which are obtained in a same batch of bio-experiments. It can be seen that the white region of each binary image roughly indicates the shapes of endothelial cells distribution, although some local errors exist in the binary images. The evaluated ratios of white region are also listed in the bottom of Fig. 5.
https://static-content.springer.com/image/art%3A10.1007%2Fs10544-011-9587-6/MediaObjects/10544_2011_9587_Fig5_HTML.gif
Fig. 5

Quantization of cells population based on a simple image process technique

However, these evaluated CDRs are not suitable to be used directly in the analysis due to the following points: (1) Apart from each cell height above the growth surface, the endothelial cells also generate a matrix over all of the culture surface. Each evaluated CDR is raised up 10 ∼ 15% to consider this thin film loading effect, for distinguishing from the case of no cell loading; (2) For the case where cells covered nearly the whole sensing domain, i.e. the 4th image in Fig. 5, the predicted value of CDR is usually much lower than the actual situation. Therefore the predicted value needs to be increased. The modified CDRs for each experimental sample are then used as the target values in neural network applications.

4.2 FRF data normalization and order-reduction

Although all experimental settings are the same at each time of the dynamic experiment, the amplitudes of every FRF measurements are varied with experimental environment and external disturbances. Consequently it is suitable to normalize the measured FRFs and scale them into a same level for comparison and analysis. On the other hand, there are multiple FRF datasets in each dynamical measurement and each FRF dataset contains a very large number of frequency spectral lines. In this work, frequency spectral lines are set to be 6,400 for each FRF and four sensory FRFs were recorded for each test. Obviously such FRF datasets are too large to directly apply into the neural network. Therefore the dimension of each FRF has to be reduced before the application of neural network.

For the FRF normalization, each spectrum is normalized with respect to the amplitude of its own first resonant mode. The reason for choosing the first resonant mode as the reference is based on the theoretical analysis results in Wu et al. (2009), which prove that the mass loading has the slightest effects on the first resonant mode of a rectangular membrane.

For the dimensionality reduction, Karhunen–Loeve (K–L) decomposition method is then used to extract the principal components on a multiple-FRFs dataset. The Karhunen–Loeve (K–L) decomposition is a useful method to create low dimensional, reduced-order models of dynamical systems (Ma et al. 2008). Assuming there are M of FRFs with N frequency in each of dynamics measurement of membrane, then this dataset forms a M × N matrix [H(ω)]M×N. The process of principal component extraction of the matrix [H(ω)] using Karhunen–Loeve (K–L) method has the following steps:
  1. 1.
    Firstly, a correlation matrix [C]M×M is created based on the FRF matrix [H(ω)]M×N.
    $$ [C]_{M\times M} = [H(\omega)]_{M\times N} [H(\omega)]^{T}_{N\times M} $$
    (3)
     
  2. 2.
    The principal components are then obtained from calculating the eigenvalues and corresponding eigenvectors of matrix [C].
    $$ [C][X] = \lambda[X]$$
    (4)
     
  3. 3.

    Finally, the M extracted eigenvalues are examined. The eigenvectors associated with these largest eigenvalues are then considered to be the principal components and be able to represent the most significant information of the original FRF dataset.

     

4.3 Dataset creation

The dynamics (FRF) of four different used membranes without any cells loading are also provided in the dataset as references. Two additional samples are also provided for the purpose of validation. Consequently, 18 different samples in total are created for training and validation of the neural network. The eigenvectors related to the largest eigenvalue of FRF dataset of each sample are extracted as the neural network input and the CDRs of every samples are calculated as the neural network targets.

4.4 Network design and training

The widely used back-propagation (BP) neural network was selected to predict cells density in this work. In the process of prediction, the principal components are extracted from FRF datasets and used as the input of BP neural network. Furthermore, the value of index AFDR of each sample provide an additional input to the neural network. As the index of AFDR has been proved to be highly related to cells distribution in last section, it can help the neural network to achieve a fast convergence and good predictions. Among the 18 samples in the dataset, the first 14 samples are used for training neural network and the left four sample are used for validation.

As the number of samples is limited, it is more suitable to design and use a simple neural network rather than a complicated one. The BP neural network used here is designed to have only one hidden layer with few neurons. Several trials with different number of hidden layer neurons were carried out to test the differences on the normalized system error. It proves that the hidden layer with five neurons produces the best performance.

The training process of BP network herein establishes an approximate function (nonlinear regression) between the inputs and targets, through iteratively adjusting the weights and biases of network to meet a setting goal (mean square error). The training parameters can affect the network convergence speed as well as the final predication accuracy. Bad parameters may lead to very slow training processes or over-fitting results. Several tests were then carried out to find optimal training parameters. The final training parameters used in this work are selected as: moment rate is 0.9, learning rate is 0.1, the max error is 0.001 and the max number of iteration is 3,000.

4.5 Prediction results

Figure 6 demonstrates the prediction results of CDR on samples of No. 15 ∼ No. 18 from the trained BP network. The prediction results match very well with the CDR values calculated from corresponding LSM images, with errors within 10%.
https://static-content.springer.com/image/art%3A10.1007%2Fs10544-011-9587-6/MediaObjects/10544_2011_9587_Fig6_HTML.gif
Fig. 6

Predicted results on the CDR of No. 15 ~ No. 18 samples using the trained BP neural network

5 Conclusion and discussion

The experiments implemented in this research have examined the biosensing performance of a micromachined rectangular silicon membrane in a normal cell culture environment. The principals of biosensing used on the micro-membrane are based on the changes of its dynamic properties caused by cell adhesion. In contrast to previous research of biosensors, the bio-experiments on each type of micro-membranes were repeated many times. Initially, the shifts of resonant frequencies were employed to analyse the experimental results. The analytical results demonstrates that the rectangular micro-membranes have capability on cell detection, under high-damping liquid conditions. Nevertheless, a fairly linear relationship of the micro-membrane sensitivity is rarely achieved. It reflects the complexity of rectangular micro-membrane in the applications of biosensing. Those results also reveal the issue of that two different entities of micro-membranes appear a certain amount differences in sensitivity, even they are of an identical type (same dimension and same boundary conditions).

Further biosensing analysis of the micro-membrane is based on the novel methodology that uses an artificial neural network with a distributed sensing scheme to estimate the adhesive cell distribution. Karhunen–Loeve (K–L) decomposition method is successfully used to reduce the dimension of measured FRF datasets. A BP neural network is trained from a set of selected experimental samples. The final predicted results on the other samples prove that this methodology can be successfully applied to identify the cell features. Significant advantages are discovered by applying this methodology in the biosensing analysis: (1) it is a robust algorithm and can repress the uncertainties in experimental measurements, comparing with using a single value as sensing parameter; (2) it is capable to eliminate individual differences of different production of micro-membrane based biosensors, consequently the efforts of calibration for each biosensor can be largely restrained; (3) it can overcome the inherent nonlinearity of sensing structure; (4) it is suitable to analyse the cells that are of very irregular shapes and non-uniform density, such as the EA.hy 926.

The work described in this paper is the first attempt of using the neural-network algorithm to perform the biosensing function of rectangular micro-membranes with a distributive sensory scheme. Much further research is required to develop more potential applications of this methodology in the field of biosensing. Although many repeated bio-experiments have been implemented in this work, the number of samples remains insufficient large for training a sophisticated neural network. Current predicted results of cell adhesion is only for the cell density spreading on the membrane sensing surface, which primarily reflects the weight information of cells. The distributive sensory data of membrane also provide the space information of the adhesive cells. Consequently, it is likely to predict the position, morphology and behaviours of living biological particles by using the proposed methodology. Such information are more useful in the biological applications than the weight information.

Acknowledgement

The authors would like to acknowledge the funding support from the EPSRC in the UK.

Copyright information

© Springer Science+Business Media, LLC 2011