Paper
Transcription
Paper
2306 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 38, NO. 12, DECEMBER 2003 A 128 128 CMOS Biosensor Array for Extracellular Recording of Neural Activity Björn Eversmann, Martin Jenkner, Franz Hofmann, Christian Paulus, Ralf Brederlow, Birgit Holzapfl, Peter Fromherz, Matthias Merz, Markus Brenner, Matthias Schreiter, Reinhard Gabl, Kurt Plehnert, Michael Steinhauser, Gerald Eckstein, Doris Schmitt-Landsiedel, and Roland Thewes Abstract—Sensor arrays are a key tool in the field of neuroscience for noninvasive recording of the activity of biological networks, such as dissociated neurons or neural tissue. A high-density sensor array complementary metal–oxide–semiconductor chip is presented with 16 K pixels, a frame rate of 2 kiloframes per second, and a pitch of 7.8 m 7.8 m for imaging of neural activity. The related circuit and system issues as well as process aspects are discussed. A mismatch-canceling calibration circuitry with current mode signal representation is used. Results from first biological experiments are presented, which prove full functionality of the chip. Index Terms—Bioelectric potentials, biological cells, biological tissues, biomedical transducers, image sensors, multi-electrode array (MEA), nervous system. I. INTRODUCTION N EURONS ARE biological cells specialized in transmission and processing of information. The elementary neural signals are action potentials, which are transient changes of the voltage drop across the cell membrane with a typical shape. The related peak-to-peak amplitudes of this transmembrane voltage approximately amount to 70 mV . Today’s standard tools used to characterize this parameter are patch pipettes [1] or microelectrodes [2]. An example setup is schematically sketched in Fig. 1(a). The patch pipette (or microelectrode) penetrates the cell membrane in order to contact the intracellular volume of the cell. The cell is situated within a grounded nutrition electrolyte. These techniques have led to great achievements in the field of neural cell monitoring, but they also suffer from a couple of disadvantages. They require a elaborate mechanical setup, which allows monitoring very few cells in parallel only. Thus, they are generally not suitable to fulfill high-throughput requirements. Moreover, long-term recording is restricted, due to the invasive type of contact, reducing the lifetime of the cell. Manuscript received May 14, 2003; revised June 5, 2003. B. Eversmann is with Infineon Technologies AG, Corporate Research, D-81730 Munich, Germany, and also with the Technical University of Munich, Munich, Germany (e-mail: [email protected]). M. Jenkner, F. Hofmann, C. Paulus, R. Brederlow, B. Holzapfl, and R. Thewes are with Infineon Technologies AG, Corporate Research, D-81730 Munich, Germany. P. Fromherz and M. Merz are with the Max Planck Institute for Biochemistry, D-82152 Martinsried, Germany. M. Brenner, M. Schreiter, R. Gabl, K. Plehnert, M. Steinhauser, and G. Eckstein are with Siemens AG, D-81730 Munich, Germany. D. Schmitt-Landsiedel is with the Technical University of Munich, D-80290 Munich, Germany. Digital Object Identifier 10.1109/JSSC.2003.819174 Fig. 1. Schematic presentation of recording techniques used to characterize neural activity. (a) Invasive patch pipette. (b) Oxide–semiconductor field-effect transistor (OSFET). Extracellular, and therefore, noninvasive, recording techniques open a way to circumvent these drawbacks. They are based on the fact that the action potentials to be considered correspond to sodium and potassium ion currents through ion channels in the cell membrane. Consequently, the idea behind these approaches is to monitor these ion currents instead of the transmembrane voltage. For this purpose, the neuron can be cultured in a nutrition electrolyte upon a planar sensor. A cleft of approximately 50 nm is formed between cell and sensor surface due to the topology of the cell membrane. Membrane currents through the ion channels located above the sensor area flow through the cleft and lead to a voltage drop due to the resistance of the electrolyte within the cleft [Fig. 1(b)]. The magnitude of this extracellular signal depends on the membrane currents and on the cleft resistance. The transmembrane currents depend on type and number of contributing ion channels, which are a function of the size of the cell and of the spatial distribution of the ion channels. The cleft resistance is a function of the contact area, of the thickness of the cleft, and of the resistance of the electrolyte. In a first-order approximation, the ion channels through the membrane can be modeled as lumped capacitance and conductance driven by the intracellular voltage with respect to action potential [3]. Membrane capacitance and cleft resistance result in a high-pass transfer function which attenuates the low frequency spectrum of the voltage signal in the cleft. Peak-to-peak values of the action-potential-induced cleft voltage are of order 100 V to 5 mV . Using an oxide–semiconductor field-effect transistor (OSFET), depicted in Fig. 1(b), as transducer, a modulation 0018-9200/03$17.00 © 2003 IEEE EVERSMANN et al.: A 128 128 CMOS BIOSENSOR ARRAY FOR EXTRACELLULAR RECORDING OF NEURAL ACTIVITY of the cleft potential is translated into a modulation of the transistor’s drain current. In the literature, this principle has been successfully demonstrated using OSFETs consisting of bulk silicon with diffused junctions and channel areas directly covered by a sensor dielectric [4]–[6]. Recording the activity of neural networks has always been a central goal in neurobiology and is gaining importance in cellbased pharma screening. However, the tools discussed so far are suitable for single-cell characterization only. Arrays of sensors are thus advantageous in all cases where several neurons must be monitored in parallel. Linear arrangements with 96 OSFETs with a pitch of 3.6 m have been suggested in the literature [7] using the same technological approach as in [4] and [5]. However, an extension of this approach to provide two-dimensional sensor arrays with similar pitch is impossible, due to the fact that only one layer (diffusion lines) can be used for interconnect purposes. Consequently, the realization of an architecture which allows row and column selection is impossible. A commercially available sensor array approach for noninvasive measurements of neural activity of neural tissue is the multielectrode array (MEA) [8]–[11]. MEAs typically consist of less than 100 planar metal electrodes on an insulating glass m and a pitch m. In case substrate with a diameter of commercially available MEAs [10], [11], measurement circuitry is usually realized by discrete off-chip components. However, from the literature, approaches for low-density arrays with active on-chip circuity are also known [12]–[14]. Considering the dimensions of neurons, which range from below 10 m for vertebrates up to 100 m for invertebrates, the need for high-density arrays is apparent to explore a maximum amount of information from cell-based biological experiments. In particular, such sensor arrays open the way for fast and statistically significant cell-based pharma screening and neural network characterization. In order to significantly increase the number and density of sensors, the integration of active circuitry is mandatory to overcome interconnect restrictions by multiplexed readout techniques, and to ensure signal integrity by signal preamplification. In this paper, a CMOS sensor array with 128 rows and 128 columns and a sensor pitch of 7.8 m is described [15]. The paper is organized as follows. In Section II, the specification of the sensor array is derived from application-related biological boundary conditions. Process and technology aspects are briefly discussed in Section III. System and circuit design are considered in Sections IV and V, respectively. Experimental results are revealed in Section VI. Finally, in Section VII, the paper is summarized and future perspectives are projected. II. SPECIFICATION The most prominent in vitro applications for high-density sensor arrays are fire-rate analysis of sparse neuron networks and imaging of brain slice activity. These two application cases are schematically sketched in Fig. 2. In the case of fire-rate analysis [Fig. 2(a)], dissociated cells cultured upon the sensor surface reestablish neural networks by the outgrowth of dendrites. For rat neurons (vertebrates) with a 2307 Fig. 2. Schematic plot describing the most prominent application cases. (a) Networks of dissociated neurons. (b) Brain slices. diameter of about 10 m, the resulting extracellular signals in the cleft have a bandwidth of approximately 10 Hz–1 kHz and an amplitude of approximately 100 V . Another suitable model system for such applications is given by snail neurons (invertebrates). Snail cells with diameters of approximately 50 m provide signals of below 3 mV . Bandwidth is 10 Hz–1 kHz. In fire-rate analysis applications, e.g., for pharma-screening purposes, it is mandatory to acquire statistically significant data. This is required to operate thousands of sensors in parallel. Slice imaging requires detecting signals from a dense organotypic layer of cells [Fig. 2(b)]. The high density of cells in the tissue translates into an increased cleft resistance, as compared with the first application case. As a result, field potentials up to 5 mV appear between tissue and sensor surface. Bandwidth requirements are the same as in the case of fire-rate monitoring. Slice imaging requires monitoring signals within a total sensor mm . area A universal sensor array platform thus requires a frame rate 2 kiloframes per second (kfps), a resolution of the cleft voltage better than 50 V, a maximum cleft voltage range of 5 mV, a m in both directions, and a total sensor area sensor pitch of at least 1 mm 1 mm. III. EXTENDED CMOS PROCESS For our sensor array, we use a CMOS process compatible adaptation of the OSFET sensor principle from Fig. 1(b). A schematic cross section of the resulting sensor is shown in Fig. 3(a). A sensor electrode coated with a biocompatible dielectric layer is connected to the polysilicon gate of a MOSFET fabricated in a standard CMOS process. Fig. 3(b) shows the corresponding lumped equivalent circuit , membrane diagram consisting of membrane conductance , and cleft resistance . Compared to the capacitance simple OSFET sensor from Fig. 1(b), the transistor gate signal and MOSFET is lowered here, since sensor capacitance gate oxide capacitance form a voltage divider. Concerning the properties of the sensor dielectric, there are biological, process related, and circuit design related requirements. During the cultivation of cells on the sensor surface for up to several weeks, the dielectric layer is in intimate contact with the cells and the nutrition electrolyte. Despite the fact that 2308 Fig. 3. Schematic cross section of the sensor principle used in this paper. (a) Cross section neuron/sensor. (b) Lumped equivalent circuit diagram. IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 38, NO. 12, DECEMBER 2003 Fig. 5. SEM photo of sensor cross section. As sensor dielectric, a 40-nm-thick stack of 10-nm TiO , 5-nm ZrO , 10-nm TiO , 5-nm ZrO , and 10-nm TiO layers is sputtered [Fig. 4(b)]. The measured effective relative dielectric constant approximately equals 45. The dielectric layers are sputtered at a temperature below 400 C to avoid damage of the metallization layers below. On the adhesion layer of the bond pads, the dielectric is removed by a reactive ion etch [Fig. 4(c)]. A 400-nm-thick Au layer is vapor deposited on the Pt pad and structured using a liftoff process [Figs. 4(d) and (e)]. A scanning electron microscope (SEM) photo of a cross section of the complete sensor with dielectric, electrode, and transistor after completion of the process is shown in Fig. 5. IV. SYSTEM SETUP AND CHIP ARCHITECTURE Fig. 4. Schematic process flow. the dielectric must be suitable for cell cultivation, i.e., biocompatible, it must be chemically inert and leakproof to avoid corrosion of the sensor electrode or of the subjacent layers of the CMOS chip. Moreover, it is obvious that the signal-to-noise ratio (SNR) of the structure in Fig. 3 improves with increasing electrode capacitance, so that a high dielectric constant and relatively low thickness of the sensor dielectric are demanded. The process flow of the extended CMOS process is schematically sketched in Fig. 4 [15], [16]. We start with a two-metallayer n-well epi CMOS process optimized for analog applicam, gate oxide thicktions with minimum gate length nm, supply voltage V, LATID-n-MOS and ness LDD-pMOS devices, poly–poly capacitors, and different types of polysilicon resistors. After this standard process (devices, two metal layers, nitride passivation, and tungsten via) is completed, the extra sensor process starts. First, the surface is planarized by a CMP step. An approximately 50-nm-thick Ti/Pt stack is vapor deposited and structured in a liftoff process. This metal stack is used to provide the sensor electrodes and the adhesion layer for the contact pads. In Fig. 6, system setup and chip architecture are depicted. The monolithically integrated sensor array consists of 128 128 pixels. The column decoder periodically selects one of the 128 columns of the pixel array. Moreover, control signals are provided by this decoder, which determine whether the pixels within the selected column are operated in the readout mode or in a calibration mode. Calibration of all pixels is performed before they are operated in readout mode, and periodically repeated after a number of readout frames (typically, 100). Circuit design aspects of calibration and readout are discussed in detail in Section V. The full frame readout rate is 2 kfps. Each of the 128 rows is connected to a separate readout amplifier. The outputs of these amplifiers are connected to the chip’s output drivers or to dummy loads via 16 8-to-1 multiplexers. The 16 output drivers force an output current into / converters arranged on an off-chip printed circuit board (PCB). Finally, the buffered output voltages of the / converters are converted by 16 analog-to-digital converters (ADCs) with an effective resolution above 8 bit, which are part of a Gage [17] PC-based measurement system. The resulting data rate amounts to 32 MS/s. A user interface program sorts and represents the acquired data. Moreover, this interface allows controlling timing of the applied control signals, the number of frames per calibration, and the power-up and EVERSMANN et al.: A 128 Fig. 6. 128 CMOS BIOSENSOR ARRAY FOR EXTRACELLULAR RECORDING OF NEURAL ACTIVITY 2309 Architecture of sensor array with complete system. power-down routines of the sensor array, by generating the corresponding control patterns. The latter are provided to the chip via magneto couplers which drive the chip’s logic interface. Specific areas of interest can also be selected. In this case, the decreased number of pixels read out translates into an increased frame rate for the considered area. Biological processes such as neural activity are strongly influenced by the environment temperature, so that this parameter must be precisely controlled and regulated. For this purpose, a resistive temperature sensor is realized on chip, and a regulation loop controls the chip’s temperature utilizing a peltier element located underneath the sensor chip. In order to control the voltage of the nutrition electrolyte, an off-chip potentiostat circuit is operated in the measurement system. For test and debugging purposes, test patterns (e.g., small AC signals) are applicable to the potentiostat’s input, which allows modulation of the electrolyte potential. An electrophysiological setup for measurement and stimulation of a single cell is added as well for reference purposes, concerning the biological activity of the selected cell. V. CIRCUIT DESIGN AND OPERATION A. Sensor Design The small signal amplitudes achieved using extracellular recording techniques, as well as the area-dependent noise contributions produced by the sensor transistor itself, demand optimizing electrode and transistor dimensions. The low-frequency noise of the sensor transistor can be reduced by increasing its area. However, the action-potential-induced modulation of its gate voltage decreases when its gate area is enlarged, due to the capacitive voltage divider formed by the sensor electrode capacitance and the transistor’s increased gate capacitance. To compensate for this effect, the sensor capacitance could be enlarged by increasing the electrode area. However, if the electrode area becomes larger than the contact area of the cell, the capacitance between the sensor electrode and the electrolyte beneath the cleft increases. Consequently, the gate voltage modulation of the sensor transistor drops again. As a consequence of these qualitative considerations, the need for an algorithmic approach to optimize transistor and electrode area is obvious. For this purpose, signal and noise are referred to the same node, the gate of the preamplifying and the SNR is investigated for different combitransistor nations of transistor and electrode area. is determined on the The gate-related signal energy basis of the known behavior of the intracellular potential during an action potential. In order to calculate the , which relates the intracellular potential transfer function to the gate voltage of , the electrical characteristics of the system consisting of membrane, cleft, electrolyte, dielectric, and electrode has to be modeled. We start with the assumption that the neuron is located in the center above the sensor electrode. Cell and electrode are assumed to be rotationally symmetric. We chose a cylindrical coordinate system whose axis coincides with the symmetry axis of cell and electrode. Moreover, uniform and time-independent distribution of conductivity and capacitance per area are assumed. In the next step, the radial variation of voltages and currents in the cleft is modeled. An equivalent circuit within the cleft [Fig. 7(a)] related to a radius increment 2310 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 38, NO. 12, DECEMBER 2003 Fig. 7. Illustration. (a) Radial incremental equivalent circuit of neuron, cleft, and electrode. (b) Corresponding matrix description. Fig. 8. (a) Schematic overview of signal path, transfer functions, and noise contributions. (b) Gate-referred equivalent schematic. is derived as a function of the distance of the center, i.e., the radius . The conductivity of the membrane within the radius , the product of the increment is , and area of the membrane within the radius increment, . Same relations hold for memthe conductivity per area, brane and electrode capacitance within the radius increment, and , with and being the capacitances of the cell membrane and electrode per area, respectively. The resistance of the electrolyte is in the cleft , and from derived from the resistance of the electrolyte, . The values of conductivity and the thickness of the cleft capacitances per area are calculated on the basis of the values , and [Fig. 3(b)] given for the lumped parameters in [18]. of sections The cleft is divided in a discrete number . Applying matrix-based according to the radius increment descriptions for the equivalent circuit of , a description of each radius increment the cleft is calculated as matrix product. With this model of the cleft and the boundary conditions concerning the grounded electrolyte outside the cleft and the gate as depicted in Fig. 7(b), the transfer function capacity of is approximated so that the gate-related signal energy of an action potential can be estimated. To model the gate-related noise, the different noise contributions have to be taken into account, as well as the consequences of readout architecture and multiplexed readout on noise propagation. Fig. 8(a) depicts the signal and noise contributions, the transfer functions, the corresponding bandwidths of the individual stages, and the multiplexing units within the signal path. has already been derived. The noise The transfer function to the gate voltage of is noise. contribution These signals are transformed via the transfer function EVERSMANN et al.: A 128 Fig. 9. 128 CMOS BIOSENSOR ARRAY FOR EXTRACELLULAR RECORDING OF NEURAL ACTIVITY 2311 Operation modes of pixel circuits. (a) Calibration. (b) Readout. into transistor drain current within bandwidth . The tranconsist of sistor’s drain current noise contributions thermal and flicker noise. The sum of these signals is connected to the input of a readout amplifier via a 1-to-128 multiplexer. In order to guarantee that the signal settles to 8 bits in the available time slot of 50% of (2 kS/s 128 columns) , the bandneeds to be at least 2.8 MHz. Noise contributions width , from the output from the readout amplifier, and , are sigdrivers and from the ADCs, nificantly smaller and can be neglected in the optimization procedure. The sum of signal and noise is amplified by the transfer , which describes signal propagation from the function readout amplifier’s input to the output driver’s input, within a . The line drivers input current is multiplexed bandwidth eight times. The input current of the output driver is amplified by the output driver, the / converter, and buffer, and the transfer is . function with required bandwidth Due to the eightfold multiplexing, the settling time is 1/8 . Consequently, of the settling time of the drain current of and needs to be at least the overall bandwidth of 22.8 MHz. At the ADC input, the weighted sum of signal and noise is sampled for one pixel with the sampling frequency of 2 kS/s. The crucial point is that the sampling frequency is sigand . nificantly smaller than the bandwidths Therefore, high-frequency noise contributes to the noise in the signal frequency domain. For the sum of noise energy contributions related to the gate, we obtain Using these approximations for the gate-related signal and noise energies, a high SNR is used as optimization goal in a MathLab-based program. For the chosen parameters, an optimum gate area of 11 m is obtained and an optimum electrode area of 16 m . Due to the layout of the pixels, this translates into a pitch of 7.8 m. Circular sensor electrodes are used with a diameter of 4.5 m. For comparison, note that typical cell diameters are between 10 and 50 m. B. Pixel Circuit The purpose of the pixel circuit is to preamplify and to buffer the extracellular signal. To ensure signal integrity within the whole array, mismatch and settling behavior have to be taken into account, as well as noise aspects. The threshold voltage mismatch of the sensor transistors is by far higher, compared with the signals to be detected. With a matching constant of 14 mV m for this process [19] and a gate area of 11 m , a one-sigma value of mV m m mV is obtained. In order to compensate for such mismatch effects, every pixel within the array is periodically calibrated so that all pixels provide the same signal-transmission properties. The calibration technique used is schematically illustrated in Fig. 9(a). The pixel in column is selected by closing the related (controlled by the column decoder). For calibration switch is closed, so that the associated sensor transistor switch is operated in a diode configuration. A constant current forces a current through the row line and the selected source . transistor is opened again and a voltage related to the caliThen, . The charge bration current is stored on the gate node of injected at the gate node of the sensor transistor during the 2312 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 38, NO. 12, DECEMBER 2003 Fig. 10. (a) Circuit diagram of complete signal path with details in (b) and (c). (b) Quasi-differential stage in readout amplifier. (c) Output driver and I /V conversion. switch-off process is approximately equal for all pixels. Moreover, the effect of the charge injection on the gate voltage is alleviated due to the large capacitance of the electrode, so that this effect can be neglected in our discussion. The procedure is performed for all rows in parallel, columns are selected subsequently. As result, all sensor transistors within a row provide the same current when selected for readout open) independently of their individual device (i.e., with parameters. In the readout mode [Fig. 9(b)], a constant voltage source is connected in parallel to current source by closing switch . Action-potential-induced changes of the drain current of a and selected sensor transistor flow through voltage source are further amplified as symbolized by the current-controlled in Fig. 9(b). current source providing the current This compensation technique has been motivated here by considering threshold voltage mismatch only. This is reasonable, since for low and medium gate voltages, the total transistor drain current mismatch is dominated by threshold voltage mismatch [19]. However, at this point, we like to mention that the applied compensation technique compensates for the variations of all sensor transistor parameters which contribute to its drain current mismatch. The discussion above also reveals that a current mode readout technique is applied. This is mandatory to meet the required pixel bandwidth, since the influence of the parasitic capaciis by far lower in curtances of row signal line and switches rent mode, as in the case of a signal representation in voltage mode. The bandwidth is now a function of the input resistance of , rather than of the properties of the pixel circuit elements. C. Complete Signal Path In Fig. 10, all subcircuits of the complete signal path are depicted. The current source in Fig. 9 is realized by transistor in the readout amplifier block. This device is used to force the calibration with respect to bias current through the selected in calibration with respect to readout mode. The transistor current is controlled by voltage , which is applied to all 128 in parallel. devices It can be shown that the chosen compensation technique reby duces the drain current mismatch of the sensor transistors . Compared to the mismatch the factor without calibration, a reduction of more than two decades is achieved for the technology and bias conditions used. Since the voltages stored on the gate nodes of the sensor transistors differ only slightly after calibration, and since the threshold voltage is negligible, mismatch of the transistors used as switches the amount of charge injection is very similar for all sensor tranare opened. sistors when switches In the readout mode, the voltage of the row signal line is kept constant by the (closed) negative feedback loop composed of and . There, the desired value amplifier and devices of the row-line potential (i.e., ) is applied to the inverting input of amplifier via an external voltage source. – in A fast pseudodifferential input stage [ is chosen Fig. 10(b)] with a small low-frequency gain to achieve stability by sufficient phase margin and the required settling time of the loop. The corresponding corner frequency of , the loop is is the parasitic capacitance of the drain-diffusion where and of the interconnect row line itself. areas of the switches EVERSMANN et al.: A 128 128 CMOS BIOSENSOR ARRAY FOR EXTRACELLULAR RECORDING OF NEURAL ACTIVITY Fig. 12. Fig. 11. 2313 Schematic measurement setup. Chip microphotograph. The difference current of and is copied and amplified by a factor of 100. Transistor is used to provide a by . This common-mode bias current through the branch branch is also calibrated before operated in readout mode. In the is closed so that is operated in calibration phase, switch diode configuration. After opening , a voltage is stored on the gate node of this transistor, which is operated as a current source . This procedure is providing the same current as transistor similar to the calibration procedure applied to sensor transistor . and is either forced The output difference current of (in parallel with the contribuinto the dummy load tions from six further readout amplifiers) or into the output stage . The drain current of drives a pMOS current mirror followed by a cascoded nMOS current mirror, with – in Fig. 10(c)]. an overall current gain of 8 [( An off-chip current source with high output resistance and low input capacitance provides a common-mode current of the output current mirror. The through the branch difference current is converted into a voltage by the resistor in the feedback loop of an off-chip discrete operational amplifier, whose output voltage is buffered again and applied to the input of the off-chip ADC. VI. EXPERIMENTAL RESULTS A microphotograph of the fabricated active sensor chip is shown in Fig. 11. The total area is 6.5 mm 5.2 mm. After sawing, the chips are die bonded with a heat-conducting adhesive into a ceramic CPGA package and bonded with aluminum wire. The cultivation chamber is glued with silicone to the package and the interspace volume with the bond wires is filled with epoxy resin. In Fig. 12, the schematic setup of sensor chip with data acquisition system, electrolyte, potentiostat, and electrophysiological reference setup is illustrated. The potentiostat is used to adjust the potential of the electrolyte to the value of a programmable . Low-pass filtering minimized off-chip reference voltage Fig. 13. (a) Photograph of measurement setup including micromanipulator for (1) microelectrode, (2) off-chip electronic components (I /V converters, buffers, magnetocouplers, temperature control, heat sink), (3) microscope, and (4) potentiostat. (b) (5) Blow-up with microelectrode, (6) ceramic package, (7) culture chamber, (8) ZIF socket, and (9) potentiostat contacts. crosstalk from off-chip noise sources. The value of is chosen so that the voltage drop across the sensor dielectric is low, in order not to avoid electrical stress. The electrolyte can be AC modulated with a test signal to allow electrical testing of the sensors of the array. The electrophysiological setup allows the creation of an intracellular contact for cell stimulation purposes and for intracellular measurements. This option is provided for reference purposes to compare extracellular and intracellular activity. Photos of the complete setup, the package, and neurons in culture are presented in Fig. 13. A. Electrical Testing For electrical tests of the chip, the electrolyte on top of the sensor array is sinusoidally modulated with an amplitude of approximately 5 mV at a frequency of 200 Hz. All sensor spots are recorded with a sampling frequency of 2 kHz. Fig. 14 depicts output signals of 12 different randomly chosen pixels within a column. The plot proves full system functionality. Mismatch, noise, nonlinearity, crosstalk, switching artifacts, and drift are sufficiently low to meet the biological requirements. B. Proof-of-Principle With Biological Samples The extracellular recording capabilities are tested with neurons from a pond snail, lymnaea stagnalis. Before cultivation, the chips are cleaned, sterilized, and coated with a cell-adhesive layer of poly-l-lysine. Then, the neurons are extracted from the animals, placed on the sensor chip, and cultured for one day. The 2314 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 38, NO. 12, DECEMBER 2003 TABLE I EXTRACELLULAR RECORDING SENSOR CHARACTERISTICS Fig. 14. Measured signals of different pixels within a column during test modulation of the electrolyte with 2 mV at 200 Hz. The figure shows the signals obtained at the ADC inputs. is recorded with the same microelectrode. Bursts of four actions potentials with an amplitude of 60 mV are obtained. The change of the intracellular potential at the beginning and the end of the current injection is a measurement artifact due to the voltage drop, which occurs at the microelectrode’s resisin Fig. 12(b)]. tance [parameter An area of 32 40 pixels of the sensor array is selected for readout with a sample rate of 8 kHz. Data of the extracellular potential recorded from one pixel beyond the neuron is shown in Fig. 16. As expected, due to the high-pass transfer function of the cleft, the extracellular signal represents the derivate of the intracellular signal. More detailed experimental results will be published in forthcoming papers with a more biology-related focus. VII. CONCLUSION AND FUTURE PERSPECTIVE Fig. 15. Snail neuron on sensor chip in culture. A biosensor array chip based on an extended 5-V 0.5- m CMOS process has been presented which allows high-resolution imaging of extracellular signals from neural cells or tissue. The chip provides 128 128 capacitive sensors with a pitch of 7.8 m 7.8 m on an area of 1 mm 1 mm. Signals to be detected are of order 100 V to 5 mV peak to peak; full-frame sample rate per pixel is 2 kS/s; maximum internal bandwidth equals 32 MHz; output data rate is 32 MS/s. The detection circuitry is based on a sensor-MOSFET mismatch-compensated current-mode technique. Data of the first biological measurements reveal successful operation of the chip. A summary of the sensor’s significant properties is given in Table I. In the next generation, the functionality of the sensor array will be increased. A bidirectional sensor/actuator interface, which allows recording as well as capacitively stimulating neural activity, will be realized and on-chip analog-to-digital conversion will be implemented. REFERENCES Fig. 16. Measured data from a snail neuron. The neuron is stimulated and monitored for reference purposes with a microelectrode. Stimulation current I , intracellular potential V , and extracellular potential V recorded with sensor array. biocompatibility of the sensor surface is proven by the growth of neurites and of the cell cultured upon the sensor chip, as shown in Fig. 15. Fig. 16 shows the result of a neuron tapped by a microelectrode and stimulated by the injection of a constant current of 0.1 nA forced into the cell for 500 ms. The intracellular voltage [1] A. Molleman, Patch Clamping. Chichester, U.K.: Wiley, 2002. [2] R. D. Purves, Microelectrode Methods for Intracellular Recording and Ionophoresis. London, U.K.: Academic, 1981. [3] P. Fromherz, C. O. Müller, and R. Weis, “Neuron transistor: Electrical transfer function measured by patch-clamp technique,” Phys. Rev. Lett., vol. 71, no. 24, pp. 4079–4082, 1993. [4] P. Bergveld, J. Wiersma, and H. Meertens, “Extracellular potential recordings by means of a field effect transistor without gate metal, called OSFET,” IEEE Trans. Biomed. Eng., vol. BME-23, pp. 136–144, Feb. 1976. [5] P. Fromherz, A. Offenhäusser, T. Vetter, and J. Weis, “A neuron-silicon junction: A Retzius cell of the leech on an insulated-gate field-effect transistor,” Science, vol. 252, no. 5010, pp. 1290–1293, 1990. EVERSMANN et al.: A 128 128 CMOS BIOSENSOR ARRAY FOR EXTRACELLULAR RECORDING OF NEURAL ACTIVITY [6] P. Fromherz, “Microelectronics meets molecular and neurobiology,” in IEDM Tech. Dig., 2001, pp. 355–358. [7] P. Fromherz, V. Kiessling, K. Kottig, and G. Zeck, “Membrane transistor with giant lipid vesicle touching a silicon chip,” Appl. Phys. A, vol. 69, pp. 571–576, 1999. [8] G. W. Gross, E. Rieske, G. W. Kreutzberg, and A. Meyer, “A new fixedarray multi-microelectrode system designed for long-term monitoring of extracellular single unit neuronal activity in vitro,” Neurosci. Lett., vol. 6, pp. 101–105, 1977. [9] J. L. Novak and B. C. Wheeler, “Multisite hippocampal slice recording and stimulation using a 32-element microelectrode array,” J. Neurosci. Methods, vol. 23, pp. 149–159, 1988. [10] MCS MEA60 System Documentation. [Online]. Available: http://www.multichannelsystems.com [11] Tensor Bioscience Brain-on-a-Chip™ Technology Documentation. [Online]. Available: http://www.tensorbio.com/html/tech.html [12] K. Najafi and K. D. Wise, “An implantable multielectrode array with on-chip signal processing,” IEEE J. Solid-State Circuits, vol. SSC-21, pp. 1035–1044, June 1986. [13] P. Irazoqui-Pastor, I. Mody, and J. W. Judy, “In vivo EEG recording using a wireless implantable neural transceiver,” in Proc. IEEE Engineering in Medicine and Biology Society Conf., 2003, pp. 622–625. [14] W. Franks, F. Heer, I. McKay, S. Taschini, R. Sunier, C. Hagleitner, A. Hierlemann, and H. Baltes, “CMOS monolithic microelectrode array for stimulation and recording of natural neural networks,” in IEEE Transducers Dig. Tech. Papers, vol. 2, 2003, pp. 963–966. [15] B. Eversmann, M. Jenkner, F. Hofmann, C. Paulus, P. Fromherz, M. Merz, M. Schreiter, R. Gabl, K. Plehnert, M. Steinhauser, G. Eckstein, D. Schmitt-Landsiedel, and R. Thewes, “A 128 128 CMOS bio-sensor array for extracellular recording of neural activity,” in IEEE Int. SolidState Circuits Conf. Dig. Tech. Papers, 2003, pp. 222–223. [16] F. Hofmann, B. Eversmann, M. Jenkner, A. Frey, M. Merz, T. Birkenmaier, P. Fromherz, M. Schreiter, R. Gabl, K. Plehnert, M. Steinhauser, G. Eckstein, and R. Thewes, “Technology aspects of a CMOS neurosensor: Back end process and packaging,” in Proc. Eur. Solid-State Device Research Conf. (ESSDERC), Sept. 2003, pp. 167–170. [17] A/D Card CS1250 and Digital Output Card CG3250 Data Sheets. [Online]. Available: http://www.gage-applied.com/ [18] M. Rentschler and P. Fromherz, “Membrane-transistor cable,” Lanmuir, vol. 14, pp. 547–551, 1998. [19] C. G. Linnenbank, W. Weber, U. Kollmer, B. Holzapfl, S. Sauter, U. Scharper, R. Brederlow, S. Cyrusian, S. Kessel, R. Heinrich, E. Hoefig, G. Knoblinger, A. Hesener, and R. Thewes, “What do matching results of medium area MOSFETs reveal for large area devices in typical analog applications?,” in Proc. Eur. Solid-State Device Research Conf. (ESSDERC), 1998, pp. 104–107. 2 2315 Franz Hofmann joined the Corporate Research Laboratories of Siemens AG in Munich in 1986. He worked on degradation effects in MOS transistors focusing on the spatial resolution of defects at the silicon/oxide interface using advanced measurement methods. Since 1989, he was involved in R&D projects concerning the on optimization of the trench capacitance of the 16 Mbit and the 64 Mbit DRAM technologies. Since 1993, he worked on advanced semiconductor devices like strained SiGe MOSFET, and novel devices and memory cells using vertical transistors. Since 1999, he has been with Infineon Technologies AG, Corporate Research, where he worked on surrounding gate transistors for DRAM generations with feature sizes below 100 nm. Since 2001, he has been active in the field of extended CMOS process development for novel biosensor purposes. He holds more than 50 patents. Christian Paulus received the diploma degree in physics at the University of Bayreuth, Bayreuth, Germany, in 1997. He is currently working toward the Ph.D. degree at the University of Magdeburg. Since 1997, he has been with Siemens AG, Corporate Technology, Munich, Germany, and Infineon Technologies AG, Corporate Research, Munich. He is working in the field of CMOS-based biosensors and advanced analog circuits. Björn Eversmann (M’03) was born in Kiel in 1972. He received the Dipl. Ing. degree in electrical engineering from the Technical University of Munich, Germany, in 2000. Since then, he worked towards the Ph.D. degree in a cooperative project between the Max-Planck-Institute Institute for Biochemistry, Martinsried, Germany, the Technical University of Munich, and Corporate Research of Infineon Technologies AG, Munich, on the development of CMOS-based bio-sensor arrays to monitor the electrical signals of nerve cells and neural tissue. Since 2003, he has been with the Research Laboratory of Infineon Technologies. Ralf Brederlow received the Dipl.-Phys. degree from the Technical University of Munich, Munich, Germany, in 1996 and the Dr.-Ing. degree from the Technical University of Berlin, Berlin, Germany, in 1999. From 1995 until 1996, he was with the Walter Schottky Institute, Garching, Germany working in the field of Si/SiGe intersubband detectors and Si/SiGe-MBE. In 1996, he joined a cooperative program between Siemens Corporate Research, Munich, and the Technical Universities of Munich and Berlin, where he worked on the characterization, modeling, and reliability of analog circuits. Since 1999, he has been with the Corporate Research Department of Infineon Technologies, Munich. There, he is a Senior Staff Engineer for technology-related circuit design. His current work includes noise-sensitive analog circuitry, polymer electronics, and sensors for biochemical applications. He has authored or coauthored some 20 technical publications. He is a member of the program committee of the International Electron Device Meeting (IEDM) and chairs the European Design Technical Working Group of the International Technology Roadmap for Semiconductors (ITRS). Dr. Brederlow is a member of the Association of German Physicists (DPG) and the German Association of Electrical Engineers (VDE). Martin Jenkner was born in Aalen, Germany, in 1968. He received the Dipl.-Phys. degree in physics from the University of Ulm, Ulm, Germany, in 1995, and the Ph.D. degree from the Technical University of Munich, Munich, Germany in 1999. From 1995 to 2000, he worked in the Department of Membrane and Neurophysics, Max Planck Institute for Biochemistry, Martinsried/Munich, Germany, on interfacing of neural cells with active semiconductor devices. Since 2001, he has been with Infineon Technologies AG, Munich, Germany, where he currently focuses on the development and applications of CMOS-based chips for recording and stimulation of neural tissue. Birgit Holzapfl graduated from the Engineering School of Siemens AG, Munich, Germany. In 1983, she joined the Siemens Research Laboratory. From 1983 to 1985 she worked on the development of surface wave filters, from 1986 to 1998 on research projects concerning GaAs hetereostructure devices and on the evaluation of the matching behavior of Silicon devices, respectively. Since 1999, she has been with Corporate Research at Infineon, where she is involved in the design of CMOS-based biosensors. 2316 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 38, NO. 12, DECEMBER 2003 Peter Fromherz received the Ph.D. degree in physical chemistry in 1969 from the University Marburg, Marburg, Germany. He was a Postdoctoral Fellow at the Max Planck Institute for Biophysical Chemistry at Goettingen, Germany. He became a full Professor for Experimental Physics at the University of Ulm, Ulm, Germany, in 1981. In 1994, he was elected as a Scientific Member of the Max Planck Society and joined the Max Planck Institute for Biochemistry in Martinsried/Munich. Currently, he is Director at the Max Planck Institute for Biochemistry and Professor for Biophysics at the Technical University Munich, Munich, Germany. Dr. Fromherz is a member of the Academy of Sciences of Berlin and Heidelberg, Germany. Matthias Merz received the degree in physics (Diplom) from the University of Ulm, Ulm, Germany in 1999. He is currently working toward the Ph.D. degree at the Max Planck Institute for Biochemistry, Martinsried, Germany, where he studies the growth of topologically defined neuronal networks that are controlled by silicon chips. Markus Brenner received the Master of Physics degree from the University of Ulm, Ulm, Germany, in 1996 and the Ph.D. degree from the Technical University of Munich, Munich, Germany, in 2001, both for work done at the Membrane and Neurophysics department of the Max Planck Institute for Biochemistry, Martinsried, on interfacing neurons with industrially made neurochip-arrays of 2048 sensors. He is currently working as an Embedded Software Architect for Siemens Corporate Technology, Munich. Kurt Plehnert received the Dipl.Ing. degree in mechanical and production engineering from the University of Dortmund, Dortmund, Germany, in 1972. Since then, he has been with Siemens Corporate Technology, Munich, Germany. He works in the field of process development with emphasis on photo lithography and thin film technology. Michael Steinhauser received the Dipl.-Chem. degree and the Dr.rer.nat. degree from the Ludwig Maximilian University of Munich, Munich, Germany, in 1980 and 1987, respectively. He joined Siemens, Munich, in October 1986, where he is working on the development of thin film modules and sensors. Currently, he is with Siemens Corporate Technology, Design to Prototype Department, where he is engaged in the development of plasmaetch, wetetch, and electroplating processes. Gerald Eckstein received the diploma degree in solid state chemistry from the University of Siegen, Siegen, Germany, in 1996, and the Ph.D. degree in materials science from the University of Erlangen-Nuremberg, Erlangen, Germany, in 2001. He joined the Corporate Technology Department of Siemens AG, Munich, Germany, in 2001, where he has been involved in back-end processes and packaging technologies of biosensors. Matthias Schreiter received the degree in electrical engineering from Dresden University of Technology, Dresden, Germany, in 1996. He joined the Corporate Research Department of Siemens, Munich, Germany, in 1996, where he worked on material and process development of functional thin films for integrated piezoelectric and pyroelectric devices. Reinhard Gabl received the M.S. and Ph.D. degrees in physics from the University of Innsbruck, Innsbruck, Austria, in 1995 and 1999, respectively. From 1999 to 2001, he was with Infineon Technologies AG, Munich, Germany, working on simulations, modeling, design, and technology development of various silicon devices, including RF-PIN, varactor and Schottky diodes, MOS transistors, and high-power RF-transistors. Since 1991, he has been with the Corporate Technology Department of Siemens AG, Munich, leading a project team specializing in functional thin films on silicon. His current interests cover the combination of silicon devices, particularly MEMS with new materials, as well as the research and development of novel MEMS sensors and sensor arrays for utilization in gas and biodetection. Doris Schmitt-Landsiedel received the Dipl.Ing. degree in electrical engineering from the Technical University of Karlsruhe, Karlsruhe, Germany, the diploma in physics from the University of Freiburg, Freiburg, Germany, and the Dr.rer.nat. degree from the Technical University of Munich, Munich, Germany. Following some research projects on semiconductor lasers and nonlinear optics, she joined the Corporate Research and Development Department of Siemens AG, Munich, in 1981. There she worked on scaling problems in MOS devices and on the design of high-speed logic and SRAM circuits. From 1989, she was Section Manager of a group of projects in future generation memory design, analog and digital CMOS and BICMOS circuits, and design-based yield analysis. Since 1996, she has been a Professor of electrical engineering and Director of the Institute for Technical Electronics at the Technical University of Munich. Her research interests are in mixed signal and low-power circuits design, failure analysis and design for manufacturability, and sensors on silicon. EVERSMANN et al.: A 128 128 CMOS BIOSENSOR ARRAY FOR EXTRACELLULAR RECORDING OF NEURAL ACTIVITY Roland Thewes was born in Marl, Germany, in 1962. He received the Dipl.-Ing. degree and the Dr.-Ing. degree in electrical engineering from the University of Dortmund, Dortmund, Germany, in 1990 and 1995, respectively. From 1990 to 1995, he worked in a cooperative program between the Siemens Research Laboratories in Munich, Germany, and the University of Dortmund in the field of hot-carrier degradation in analog CMOS circuits. Since 1994, he has been with the Research Laboratories of Siemens AG and Infineon Technologies, where he was active in the design of nonvolatile memories and in the field of reliability and yield of analog CMOS circuits. From 1997 to 1999, he managed projects in the fields of design for manufacturability, reliability, analog device performance, and analog circuit design. Since 2000, he is responsible for the Laboratory on Few Electron Circuits of Corporate Research of Infineon Technologies. His current interests include electronic biosensors on CMOS, device physics-related circuit design, and advanced analog CMOS circuit design. He has authored or coauthored some 80 publications. Dr. Thewes is a member of the German Association of Electrical Engineers (VDE). He served as a member of the technical program committees of the International Electron Device Meeting (IEDM) and of the European Symposium on Reliability of Electron Devices, Failure Physics and Analysis (ESREF). He is a member of the technical program committees of the IEEE International Solid-State Circuits Conference (ISSCC), of the International Reliability Physics Symposium (IRPS), and of the European Solid State Device Research Conference (ESSDERC). In February 2003, he received the ISSCC Jack Raper Outstanding Technology Directions Paper Award for his work on fully electronic DNA sensor chips presented at the ISSCC 2002. 2317