Appendix 3

More Than You Ever Really Wanted to Know About

Charge-Coupled Devices
James B. Pawley

INTRODUCTION are kept electrically separate from their neighbors by additional

layers of SiO2. Every third stripe is connected together to form
The electronic structure of crystalline Si is such that electromag- three sets of interdigitating strips that we will refer to as Phases 1,
netic waves having the energy of light photons (1.75–3.0 electron 2 and 3 (f1, f2, f3, Fig. A3.1). Taken together, all these phases con-
volts) can be absorbed to produce one free or “conduction” elec- stitute the vertical register (VR) and, after the assembly has been
tron. If an image is focused onto a Si surface, the number of the exposed to a pattern of light, they are used to transfer the photo-
photoelectrons (PE) produced at each location over the surface is induced charge pattern downwards, one line at a time. The pixels
proportional to the local light intensity. Clearly, all that is needed along each line are separated from each other by vertical strips of
to create an image sensor is a method for rapidly converting the positively doped material injected into the Si. These positive
local PE concentration into an electronic signal. After almost 40 “channel blocks” create fields that prevent charge from diffusing
years of NASA and DOD funding, the slow-scan, scientific-grade, sideways without reducing the active area of the sensor.
charge-coupled device (CCD) camera is now an almost perfect Any photon that passes through the stripes and the SiO2, is
solution to this problem. absorbed in the Si, producing a PE. If a small positive voltage (~15
Success in modern biological light microscopy depends to an volts) is applied to the f1 electrodes, any PE produced nearby will
ever-increasing extent on the performance of CCD cameras. be attracted to a location just below the nearest f1 strip (Fig. A3.2).
Because such cameras differ widely in their capabilities and are As additional PEs are produced, they form a small cloud of PEs
also items that most biologists buy separately, rather than as part referred to as a charge packet. The number of PEs in the packet is
of a system, some knowledge of their operation may be useful to proportional to the local light intensity times the exposure period
those practicing biologists who have not yet found it necessary to and the problem now is to convey this packet to some location
be particularly interested in “electronics.” Although the basics of where its size can be measured, and to do this without changing it
CCD operation are described in many other chapters (particularly, or losing track of the location from which it was collected. This
Chapters 4, 10 and 12) this Appendix describes the operating will be achieved by using the overlying electrodes to drag the
principles of these devices in greater detail and also discusses the charge packet around in an orderly way until it is deposited at the
ways that they “don’t work as planned.” It then covers the opera- readout node of the charge amplifier.
tion of the electron-multiplier CCD (EM-CCD), a new variant
that reduces the read noise almost to zero, although at the cost Charge Coupling
of reduced effective quantum efficiency (QEeff).1 The second
section, How to choose a CCD, is a review of CCD specifications The dragging mechanism operates in the following way: First f2
with comments on the relevance of each in fluorescence is also made positive so that the cloud diffuses to fill the area
microscopy. underneath both f1 and f2. Then f1 is made zero, forcing the packet
to concentrate under f2 alone (Fig A3.2).
So far, these 3 steps have succeeded in moving the charge
PART I: HOW CHARGE-COUPLED packets that were originally under each of the f1 electrodes down-
DEVICES WORK wards by one phase or 1/3 of a “line” in the x-y raster. If this
sequence is now repeated, but between f2 and f3 and then again
The first step is to imagine a rectangular area of the Si surface as between f3 and the f1 belonging to the next triplet of strips, packets
being divided into rows and columns, or more usually, lines and will have moved down by the one entire raster line. PEs created
pixels. Each pixel is between 4 ¥ 4 mm and about 24 ¥ 24 mm in within a particular pixel of each horizontal stripe remain confined
size and the location of any pixel of the surface can be defined in by the channel stops as they are transferred to the next line below.
terms of it being x pixels from the left side, on line y. A pixel of the image is therefore defined as the area under a
To construct an actual system like this, start with a smooth Si triplet of overlying, vertical charge-transfer electrodes and
surface; cover it with a thin, transparent, insulating layer of SiO2; between two neighboring channel blocks. The pixels on scientific
deposit onto the SiO2, a pattern of horizontal strips, made out of a CCDs, are usually square, 4 to 30 mm on a side while those on
transparent conductor called amorphous silicon (or poly-silicon), commercial, video CCDs are likely to be wider than they are high,
so that the strips cover the entire image sensor area. Although, to conform with the reduced horizontal resolution of commercial
viewed from the top, these strips partially overlap each other, they video standards. Only square pixels can be conveniently displayed
in a truly digital manner. Larger pixels have more leakage current
(dark-current), but are also able to store more charge per pixel (see
1
This loss can be avoided if the system is used in photon-conting mode. Blooming, below).

James B. Pawley • University of Wisconsin, Madison, Wisconsin 53706

918 Handbook of Biological Confocal Microscopy, Third Edition, edited by James B. Pawley, Springer Science+Business Media, LLC, New York, 2006.
 More Than You Ever Really Wanted to Know About Charge-Coupled Devices • Appendix 3 919
BASIC CCD ARRAY
Vertical
One pixel Control electrodes
phase
Φ1
Φ2
Φ3
Drive pulse
connections

Image Channel stop

section
Parallel channel
Serial channel

Readout
section
Output

Readout node

Φ4 Φ5 Φ6

Horizontal phase
Drive pulse connections

FIGURE A3.1. Layout of CCD array, viewed en face.

FIGURE A3.2. Charge coupling: Three stages in the

process of moving a charge packet initially beneath
phase 1 (A), so that it first spreads to be also under phase
2 (B) and finally is confined to entirely under phase 2
(C). These 3 steps must be repeated 3 times before the
charge packet has been moved downwards (or in the
diagram, to the right) by one line of the CCD array. C
920 Appendix 3 • J.B. Pawley

At the bottom of the sensor, an entire line of charge packets is ferred past bright features in the image, producing vertical streak-
simultaneously transferred to the adjacent pixels of the horizontal ing. This problem is more important when the exposure time is
register (HR, also sometimes called a shift register). Like the VR, short relative to the readout time.
the HR is composed of a system of overlying poly-Si electrodes In frame transfer readout, at the end of the exposure, the
and channel stops. Each column of pixels in the VR is eventually entire charge pattern is rapidly (0.1–3 ms) transferred by charge-
transferred directly into the same specific pixel on the HR. The coupling to a second 2D storage array. The storage array is the
three phases of the HR (f4, f5, f6) work exactly like those in the same size as the sensor array and is located next to it but it is phys-
VR, except that they must cycle at a much faster rate because the ically masked with evaporated metal to shield it from light. The
entire HR must be emptied before the next line of packets is trans- charge pattern is then read out from the storage array while the
ferred down from the bottom line of the VR. In other words, in the sensor array collects a new image. Because vertical transfer can be
time between one complete line-transfer cycle of the VR and the much faster if the charge packets do not have to be read out, this
next, the horizontal register must cycle as many times as there are system reduces streaking by up to 1000¥ but does not eliminate it
pixels in each line. and the need for a storage register reduces the fraction of the Si
At the right-hand end of the HR is a charge amplifier that mea- surface area that can be used for sensing by 50%.
sures the charge in each packet as it is transferred into it from the In interline transfer, the masked storage cells are interlaced
last pixel of the HR. The first pixel to be read out is that on the between the sensor cells (i.e., each pixel is divided into sense and
extreme right-hand side of the bottom line. The last pixel will be read areas). After exposure, all charge packets can be moved to the
that on the left side of the top line.2 readout array in less than a microsecond. This ability can be used
The entire charge-transfer process has the effect of coding as an electronic shutter to eliminate vertical smearing but, because
position as time. If we digitize the signals from the charge ampli- at least half of the area of each sensor must be masked, and any
fier, and store the resulting numbers in a video memory, we will light striking a masked area is lost, the “fill factor” of the sensor is
be able to see a representation of the light intensity pattern strik- reduced, proportionately decreasing QEeff. A solution to the “fill-
ing the sensor on any monitor attached to this video memory. Alter- factor” dilemma is to incorporate an array of microlenses, aligned
natively, as long as the dimensions of the CCD array match those so that there is one above every pixel. With such a system, most of
of some video standard, such as NTSC or PAL, the time sequence the light striking any pixel will be focused onto the unmasked area.3
of charge-packet readout voltages can be smoothed and, with the Although microlenses restore the QEeff somewhat, the full-well
addition of synch pulses, turned into an analog video signal. While signal possible is still limited by the smaller sensitive area.
this latter process is often convenient, it is a poor plan if the analog
signal must then be re-digitized. The necessity to digitize twice can
reduce the effective horizontal resolution of the CCD sensor by WHAT COULD GO WRONG?
about a factor of 2 and because the process is AC coupled, photo-
metric accuracy is severely compromised. When I first heard the CCD story, it struck me as pretty prepos-
It is important to understand the relationship between the terous! How could you get all the correct voltages (9 different
charge-transfer electrodes and the charge packet. The electrodes voltage combinations per pixel shift, ~3.6 million for each TV
do not somehow “connect to” the charge packet, and “conduct” it frame, 108 million/s for video rate!) to the right charge-transfer
to the amplifier. Such a process would be subject to resistive losses, electrodes at the correct times? How could you get all of the charge
charge would be lost and a lot of “wires” would be needed. The in a packet to stay together during a transfer? Wouldn’t Poisson
charge-coupling process is better thought of in terms of a ball statistics apply, making even one transfer imprecise and the 2000
bearing “dragged” over the surface of a loose blanket by moving transfers needed to read out the top, right pixel of a 1000 ¥ 1000
a cooking pot around underneath the blanket. The weight of the pixel array impossibly inaccurate? How long would the PEs stay
ball and the lip of the pot create a dimple and gravity keeps the free to be dragged around the lattice? Wouldn’t the charge packets
ball in the dimple as the pot is moved. The voltage on the charge- decay with time?
transfer electrode creates an electronic “dimple.” Changing the In fact, many of these problems did occur, but remedies to most
voltages on nearby electrodes moves the dimple. In this way, have now been devised. The difference between a $300 commer-
groups of charged particles (electrons) can be pushed around cial CCD camera and a $65,000, top-of-the-line scientific CCD can
without actually “touching” or losing them. often be measured in terms of how many of these remedies have
been implemented. Therefore, it is worthwhile trying to understand
some of them so that one can buy what one needs. The following
Readout Methods discussion will define and discuss some of the more important
There are three distinct methods for reading out the charge pattern CCD technical specifications.
of a CCD: full-frame, full-frame transfer and interline transfer
(Fig. A3.4). Most early scientific CCDs used the first method,
which operates as has just been described. Although full-frame
Quantum Efficiency
readout provides the largest sensitive area for a given area of Quantum efficiency is the ratio of the number of impinging
silicon, the lowest level of readout noise and the greatest photo- photons to the number of PEs produced.4 Any photon with energy
metric accuracy, it also has some disadvantages. One cannot both in the range of 1–100 eV striking crystalline Si has a very high
collect and read out signal at the same time. Unless some sort of probability of producing a PE. However, reflections and absorp-
shutter is used to prevent light from striking the sensor during ver- tion by the overlying polysilicon electrodes,5 reduce the QE of
tical transfer, signal will be added to any packets that are trans-
3
This occurs only as long as the initial angle of incidence is near to normal, a
condition met when CCDs are used for light microscopy.
2 4
This may seem backwards until one remembers that any image of the real In the visible range, each absorbed photon makes only one PE.
5
world is usually focused onto the CCD by a single, converging lens, a process Kodak had pioneered the use of charge transfer electrodes made out of In and
that always inverts the image. Sn oxides that scatter less light than do those made of poly-Si.
 More Than You Ever Really Wanted to Know About Charge-Coupled Devices • Appendix 3 921

FIGURE A3.3. Four CCD readout patterns: Full-frame,

frame-transfer, interline transfer and gain register (EM-
CCD).

front-illuminated CCDs especially in the blue end of the spectrum. amplifier noise. One can also see that at readout speeds higher than
To reduce this effect, some UV-enhanced sensors are coated with 1 MHz (or 1 second to read out a 1024 ¥ 1024 CCD), the read
fluorescent plastics, which absorb in the blue and emit at longer noise increases with the square root of the read speed.
wavelengths. Others have their backs etched away and are turned
over to permit the illumination to reach the light-sensitive area
from the back side.6 Figure A3.4 shows the intrinsic QE of differ-
Charge Loss
ent types of CCD (not Qeff, which would take into account the light The lifetime of a PE (before it drops back into the ground state)
lost if some of the sensor is covered by charge storage areas). The depends on the purity and crystalline perfection of the Si and on
effective QE can usually only be determined by actual measure- other factors such as temperature. Generally it is long enough that
ment or by very careful evaluation of the published specifications little charge is lost during the exposure times commonly used in
(QEeffective = QEintrinsic ¥ fill factor). fluorescence microscopy. If necessary, it can be increased by
cooling the detector, something often done to reduce dark charge.
Edge Effects
In early CCDs, PEs were often “lost” in the crystalline imperfec-
Leakage or “Dark Charge”
tions that are always present at the Si/SiO2 junction. To avoid this, Dark charge is the charge that leaks into a pixel during the expo-
ion implantation is now used to make an N-doped, sub-surface sure time in the absence of light. It can be thought of as the dark
layer called the buried channel about 1 mm below this surface (Fig. current7 deposited into one pixel. Many processes other than
A3.2). This channel attracts the free PEs, keeping them away from photon absorption can add PE to the charge packet. The magni-
the edge of the Si crystal. Any serious CCDs will have a buried tude of this dark charge depends on the length of the exposure, and
channel but the need for ion-implantation keeps CCD chip prices is substantially reduced by cooling. The rule of thumb is that for
high! Figure A3.5 shows the readout noise, in root-mean squared every 8°C of cooling, the dark charge is halved. As noted above,
(RMS) electrons/pixel, for surface and buried-channel CCDs dark charge is principally a problem because it produces Poisson
having two different pixel sizes. From this you can see that small noise equal to the square root of its magnitude, and if this is
pixels (here ~5.5 ¥ 5.5 mm) have lower read noise than larger ones left unchecked, it can significantly increase the noise floor of the
(~17 ¥ 17 mm), mostly because the larger ones have higher capac- CCD.
itance and capacitance is the most important parameter of read- Since ~1987, a process called multipinned phasing (MPP) has
been available to reduce dark charge build-up by about a factor of
1000, making it immeasurable in exposures up to a minute or so.
6
Back-illuminated CCDs have to be thinned to 7–10 mm so that conduction This feature should be specified if one expects to use exposures
electrons created near what would have been the back surface can respond to longer than a few seconds without deep-cooling.
the fields created by the buried channel and the CC electrodes. Thinning
increases cost and also reduces QE at longer wavelength where the absorp-
tion distance of the photons becomes comparable with the actual thickness.
7
Back-illuminated CCDs are also more expensive because it is difficult to A current is a flow of charge measured in charge/time. The unit of charge is
create electrical contacts with electrodes, etc., that are now on the bottom side the Coulomb (c). The unit of current is the Ampere (A). One Amp represents
of the chip. a flow of one Coulomb/s or 6.16 ¥ 1018 electrons/s.
922 Appendix 3 • J.B. Pawley

FIGURE A3.4. Intrinsic QE as a function of wave-

length for a front-illuminated CCD (blue), a visible-
enhanced, back-illuminated CCD (green) and a
UV-enhanced CCD (red).

It should also be remembered that, while dark charge is never the number of electrons/pixel it represents. CCDs should always
good, its average value can be measured and subtracted on a pixel- be operated such that the noise on the dark charge is less than the
by-pixel basis, by subtracting a “dark image” from each recorded readout noise. On conventional CCDs this condition can usually
image as part of flat-fielding. However, because, by definition be met quite easily by slightly cooling the sensor (0°C or about
“dark” images contain very few photons/pixel, they have relatively -20°C from ambient). The use of lower temperatures is compli-
high Poisson noise and low S/N. Therefore, a number of such cated by the risk of condensing atmospheric water, a process that
images must be averaged to produce a correction mask that is sta- can be avoided only by enclosing the sensor in a vacuum chamber.
tistically defined well enough that subtracting it from the data does Generally, a vacuum-hermetic enclosure, combined with good
not substantially increase the noise present in the final, corrected outgassing prevention, carries with it the significant benefits of
image. more effective cooling, long-term protection of the sensor from
This is not a problem when there are many counts in each pixel moisture and other degrading organic condensates as well as
because the subtractive process of dark-charge normalization the prevention of front-window fogging. At video rate, where
involves a change that is small compared with the intrinsic noise exposures are short, dark charge is only a problem when the
present in a large signal. It can be a problem when the black mask readout noise is reduced to <1 e/pixel, as it is when an “electron-
image is subtracted from a faint image that also contains only a multiplier” (EM) charge amplifier is used (see below and also
few counts/pixel. Chapters 4 and 10). In EM-CCDs the read noise is so low that dark
What cannot be removed by flat-fielding is the Poisson noise current becomes the main source of noise and cooling to -80°C
associated with the dark charge. This is equal to the square root of becomes necessary.
Readout noise, electrons, rms
100

Measured
50
Surface channel
30 A = 300 µm2

20

Surface channel
10 A = 30 µm2
Buried channel (T<200K)
A = 300 µm2
5

3
Buried channel (T<200K)
2 A = 30 µm2 FIGURE A3.5. CCD field effect transistor (FET)
noise as a function of pixel dwell time for large and
small pixels and when using buried channel vs. surface
1
channels. Smaller pixels have less read noise because
10 ns 100 ns 1 µs 10 µs 100 µs Clamp-to-sample time they have less capacitance. Buried channels have almost
100 MHz 10 MHz 1 MHz 100 kHz 10 kHz Pixel readout rate 10¥ less read noise than surface channels.
 More Than You Ever Really Wanted to Know About Charge-Coupled Devices • Appendix 3 923

TABLE A3.1. Typical Performance of Various Types of CCD Cameras. The “Sensitivity” Column Is a Reasonable Estimate of
the Relative Suitability of the Camera for Detecting Very Faint Signals. It Spans a Very Large Range of Performance!
All CCDs are not equal
Type Grade QE % (effective) Noise (e/pix) Sensitivity (relative) Bit depth Dynamic Range
Video commercial color 10 200 1 10 1,000
monochrome 20 200 2 10 1,000
Digital 1 Mhz, color 15 50 12 12 4,000
1 Mhz, mono 30 50 24 12 4,000
Back. Illum/ 90 5 720 15 40,000
slow-scan
LLL-CCD 45 0.1 (18,000)* ?* 200,000
(EMCCD)
* Because the gain of the electron multiplier amplifier is unknown and large, it is not simple to measure, or even define, the sensitivity and bit depth of the EM-CCDs.

Blooming scientific applications), operate with much (100¥?) less perfection.

In microscopy today, we find CCDs that span this range of per-
As more photons are absorbed, the charge packet clustered around formance (Table A3.1).
the buried channel grows and mutual repulsion between these All CCDs are not equal!!
electrons renders the field imposed by the charge-transfer electrode
ever less successful in keeping the packet together. The maximum
charge packet that can be stored without it overflowing into nearby CHARGE AMPLIFIERS
pixels can be estimated by multiplying the pixel area (in square
micrometers) by 600 PE/pixel (i.e., 27 k PEs for a 6.7 ¥ 6.7 mm So far, I have described an image sensor in which up to 90% of
pixel, 540 k PEs for a 30 ¥ 30 mm pixel). This overflow problem is the impinging photons make free PEs and explained how the
referred to as “blooming” and, in CCDs for the home-video market charge packets that result from many photons hitting a given pixel
it is limited by the presence of an n-layer, deeper in the Si. When can be conveyed to the charge amplifier, in a time-labeled manner
the charge packet gets too big, mutual repulsion between the PEs and almost without change. Clearly the performance of the entire
forces some of them into this overflow layer, through which they image detector will depend crucially on the capabilities of this
are conducted to ground. amplifier.
While this anti-blooming feature is convenient for removing
the effects of the specular reflections found in images from
everyday life,8 it is not incorporated into many full-frame or frame- What Is a Charge Amplifier?
transfer scientific CCDs because it reduces QE for long- Although most scientists have had some exposure to electronic cir-
wavelength light. As this light penetrates farther into the Si crystal cuits that amplify an input voltage or current, they may be less
before being absorbed, much of it reaches the overflow layer where familiar with the operation of the type of charge amplifier found
any PEs produced are lost. in a CCD. The following outline is presented to enable the reader
to understand enough about the process to appreciate some of the
Incomplete Charge Transfer important differences between the various types of CCD.
Because of the pulsatile nature of the CCD charge delivery
Sometimes, an imperfection in the Si will produce a pixel that system, the optimal way to measure charge packet size is to deposit
“leaks” charge. Charge deposited into, or transferred through, this it into a (very) small capacitor (the “read node”) and then measure
pixel will be lost, producing a dark vertical line above it. In addi- the voltage on this capacitor with a high impedance amplifier. As
tion, if one pushes the pixel clock too fast, some PE in the packet a field-effect transistor (FET) has an almost-infinite input imped-
will not move fast enough and they will be left behind. In general ance, it is ideal for this purpose and in fact, its existence makes
however, on a slow-scanned, scientific CCD, fewer than 5 PEs out charge-amplification possible.
of a million are lost (or gained) in each, slower, vertical transfer There are two basic types of conventional CCD readout
and only 50 (0.005%) are lost during each, faster horizontal trans- amplifier, non-destructive and destructive.10 Both employ FET
fer. In such devices, the main noise term is Poisson noise for any amplifiers.
signal level above ~20 PE/pixel,9 and it seems hard to imagine Non-destructive (“skipper”) amplifiers use an FET with a
doing much better than this except for signal levels <16 PE/pixel. “floating gate” to sense the size of a charge packet by responding
On the other hand, it is also true that the vast majority of CCDs to the moving field that is produced as the packet is transferred
made (those for camcorders, surveillance cameras and even many along a charge-coupled register. Because the charge packet itself
is not affected by this process, the process can be repeated hun-
dreds or even thousands of times. If the results of all these mea-
8
Features such as the image of the sun reflecting off a shiny automobile can
surements are averaged, very low readout noise levels (>±1
be over 1,000¥ brighter than the rest of the scene. Fortunately, such extremely electron/pixel) can be obtained, but at the cost of a substantial
bright features are seldom found in microscopic images unless a crystal of
fluorescent dye occurs in the field of view.
9
This calculation assumes that the read noise is 4 e/pixel, and this will be less
10
than the Poisson Noise for any signal >16 PE. However, as many CCDs used The “electron-multiplier” amplifiers mentioned previously, act essentially as
in microscopy have >4 e/pixel of noise, this cut-off point should not be con- pre-amps to the conventional FET amps described here. They will be covered
sidered inflexible. later in this Appendix.
924 Appendix 3 • J.B. Pawley

Reset would, itself, produce a random electronic noise signal larger than
trigger this, and electronic noise increases with readout speed, read-node
VReset capacitance and, to a lesser extent, temperature.
Supply voltage The success of the CCD in overcoming this limitation depends
on two factors:
Reset
transistor
• The extremely small capacitance of the read node compared
to that of any other photosensor such as a photodiode.
Readout transistor, Gain G
∆V = Q • Special measurement techniques such as correlated double-
sampling
Charge Cn
input Output, G ∆V = GQ Clearly there are a lot of tricks to making the perfect CCD
Cn amplifier and not all CCDs employ them. Table A3.1 lists typical
Q Load performance for a variety of common camera types.

CCD Register Substrate and ground

NOISE SOURCES IN THE
CHARGE-COUPLED DEVICE
Readout node
capacitance
Fixed Pattern Noise
FIGURE A3.6. Destructive read-out amplifier for a CCD chip. When exposed to a uniform level of illumination, some pixels in
a CCD array will collect more charge than others because of small
differences in their geometry or their electrical properties. Conse-
(¥100 or ¥1,000) increase in readout time and logical complexity. quently, it can be necessary to use stored measurements of the rel-
This approach might make sense on a Mars probe but it has not ative sensitivity of each pixel to normalize, or “flat field,” the final
been used in microscopy to my knowledge. dataset on a pixel-by-pixel basis. This is accomplished by first
Destructive readout amplifiers are more common, probably recording an image of a featureless “white” field. This is often
because they can operate more rapidly (Fig. A3.6). As imple- approximated by a brightfield transmission image with no speci-
mented in a scientific CCD, the charge amplifier consists of the men, a process that will also record “inhomogeneity,” or mottle,
following components: in the optical system. Differences in gain between pixels are
evident as visible as nonuniformities in the digital signal stored in
• overlying charge transfer electrodes to drag the next charge the memory and these are used to derive multiplicative correction
packet into the “read node”
coefficients.12
• the read node itself: a 0.03–0.1 picofarad capacitor Unfortunately, one can only preserve the high precision of the
• the sense FET CCD output if the coefficient used to normalize each pixel is
• the reset FET equally precise. In any event, these correction coefficients vary
In operation, fields from the overlying charge-coupling elec- with both the photon wavelength and the angle at which the light
trodes force a charge packet into the read-node capacitor, creating passes through the polysilicon electrodes on its way to the buried
a voltage, Vc, that is proportional to the amount of charge in the channel. This, in turn, depends on the details of the precise optical
packet. This voltage is sensed by the sense FET and the output is path in operation when an image is recorded and may even change
passed, via additional amplifiers, to the analog-to-digital converter with microscope focus! As the intrinsic noise of a pixel holding
(ADC) where the signal is converted into a digital number. Finally, 360 k, PEs is only ±600 electrons or 0.16%, pixel-to-pixel nor-
just before the next charge packet is coupled into the read node, a malization for changes in sensor gain is seldom perfectly effective
reset FET discharges the capacitor, forcing Vc to zero, and allow- and consequently there is usually some level of “Fixed-pattern
ing the read-FET to sense it again. noise” superimposed on the final data.
In addition, the “white” image that must be used for pixel-level
sensitivity normalization is itself subject to intrinsic noise (±600
FET Amplifier Performance
electrons for a signal from a pixel with a full well charge of 360 k
The signal current (signal charge/s) coming from a CCD sensor is electrons) and so multiplicative normalization may actually add
very small. Suppose that there were, on average, 400 PE in every some noise to the raw, uncorrected signal! Fortunately, if the white
pixel of a 512 ¥ 512 pixel sensor.11 Reading this out in one second image can be defined by a multi-frame averages of several, nearly
would constitute a current of only 10-11 Amps. The current through full-well “white” images, this normalization noise should only be
the bulb in a home flashlight is 1010 times more. A very good con- noticeable when the image data to be corrected is similarly noise
ventional electronic amplifier designed to amplify this current free. Without details of the signal levels present or the optical
system in use, it is difficult to estimate the magnitude of normal-
ization noise but it will be comparatively less important for images
11
Although this number may seem small, it is actually quite high compared to
of faint objects containing few counts/pixel because these mea-
some uses in biological confocal microscopy. Many authors have found that surements are themselves less precise.
in normal” use, a single-beam confocal microscope used to image a fairly
faint stain will count 4–8 PE/pixel in bright areas of the image. Allowing that
the effective QR of a good CCD will be ~10¥ higher than that of the photo-
12
multiplier tube used in the confocal microscope, this makes the expected These correction coefficients are small and only needed when operating on
peak CCD signal in an image from a disk-scanning confocal microscope only images involving large numbers of photons (and consequently having rela-
40–80 PE/pixel. tively low Poisson noise and good S/N).
 More Than You Ever Really Wanted to Know About Charge-Coupled Devices • Appendix 3 925

It should be also noted that the vignetting and “mottle” visible Where Is Zero?
in images characteristic of video-enhanced contrast microscopy
will produce small intensity errors in the data obtained by both A final important feature of the CCD readout is that, compared to
widefield and confocal. However, this noise term will be more the photomultiplier tube (PMT), it is relatively difficult to deter-
noticeable in widefield where more photons are used and hence the mine the exact output signal level that corresponds to a zero-light
precision of the data is greater. Mottle is produced by dirt and signal. A properly operated PMT never records negative counts.
surface imperfections on any optical components that are not However, as the electronic readout noise of a cooled-CCD is an
located exactly at aperture planes, as well as by non-uniformities RMS function with both positive and negative excursions, there
in the image sensor. Fortunately, to the extent that it is stable with will be some pixels that measure lower than the mean value of the
time, mottle will be removed by the flat-field correction for CCD zero-light pixel intensity distribution.
sensitivity just discussed. To ensure that no data is “lost,” scientific CCDs are usually set
What will not be removed is any change in signal caused by up so that the zero-light signal is stored to be a few tens of digital
stray light (room light, light that goes through filters designed to units (ADU) above zero. A histogram of numbers stored from a
remove it, etc.). The simplest test of any CCD set-up is to record “black” image will show a Gaussian-like peak centered at the
an image of “nothing” (i.e., room dark, no excitation, no specimen offset and with a half-width equal to 2¥ the RMS read noise (see
etc.). Then do the same with 100¥ longer exposure time with the Fig. 4.20). This offset makes it more difficult to apply the gain and
room lights at your normal operating level. Now adjust the display offset normalization procedures to images that record only a few
look-up tables so that you can “see the noise” in both the images detected photons in each pixel, a factor that will become more
on the screen. Although the only difference between the two important as CCDs are increasingly used to image living cells that
images should be increased dark noise in the image with the longer cannot tolerate intense illumination and which therefore produce
exposure, this is seldom the case. substantially lower signal levels.

Noise from the Charge Amplifier A NEW IDEA: THE GAIN REGISTER AMPLIFIER!!
Noise is generated by both the readout and the reset FETs in the
charge amplifier. Noise generated in the readout FET reaches the Early in 2002, a new type of readout amplifier was introduced by
ADC directly. If thermal noise in the reset FET prevents it from Texas Instruments (Houston, TX) and E2V Technologies (Chelms-
completely discharging the read-node capacitance, it produces a ford, UK). As only E2V makes back-illuminated sensors, I will
random offset at each pixel (i.e., the read-node voltage is not reset describe their system but both work along similar lines. E2V orig-
exactly to zero). This is referred to as Reset Noise and has the inally referred to their device as the “gain register” and its purpose
effect that the dark charge seems to vary from pixel to pixel. For- is to amplify the size of the charge packet before it arrives at the
tunately, Reset noise can be almost eliminated by employing the read node. Although the term gain register has recently been
technique of Correlated Double-sampling (CDS) in the readout replaced by the term “electron multiplier”, it is important to
amplifier. In CDS, the circuitry of the charge-to-voltage amplifier remember that these new detectors work on a completely different
is modified so that the output is proportional to the difference principle from that employed in intensified-CCDs.
between the value of Vc just after the reset pulse and its value after The gain register superficially resembles an additional HR,
the next charge packet has been inserted. with two important differences:
Although CDS essentially eliminates the effect of reset noise,
it also distorts the noise spectrum. On the one hand, this distortion • There are 4 phases rather than the usual 3 and the new
phase consists of a grounded electrode located between f1
has the beneficial effect of converting the low frequency, 1/F noise
and f2.
from the FET into broadband noise which is more easily treated
theoretically and which is less visually distracting than the short, • The charge transfer voltage on f2, is now variable, between
+35 and +40 volts rather than the usual +15 volts.
horizontal flashes characteristic of 1/F noise.13 On the other hand,
it means that the input to the ADC must be carefully frequency- As a result, when f2 is excited, there exists a high electric field
filtered. This filtering can be implemented either by employing RC between it and the grounded electrode. The high field accelerates
circuits or by using dual-slope integration (DSI) in the ADC itself. the electrons in the charge packets more rapidly as they pass from
If there are large intensity variations between neighboring pixels, f1 to f2 with the result that each PE has a small (but finite; usually
the use of RC circuits will effectively compromise the large in the range of 0.5% to 1.5%) chance of colliding with a lattice
dynamic range of the CCD. Therefore, ADCs using DSI are electron and knocking it into the conduction band (Fig. A3.7).
employed on most slow-scan scientific, cooled-CCDs. Assuming the 1% gain figure, this means that for every 100 PE in
The fact that CDS and, in particular, DSI work best at low the packet, on average one of these will become two electrons
readout speeds is a final reason why most scientific CCDs operate before it reaches the space under f2. Although this seems like a
best at relatively low readout speeds (Fig. A3.5). The other two trivial improvement, after it has been repeated as part of the 400
reasons are improved charge transfer efficiency and the reduction to 590 transfers in the gain register, a total average gain of hun-
in broadband electronic noise from the FETs (noted above.) dreds or even thousands is possible. If the voltage on f2 is reduced
to normal levels, the sensor operates as a normal CCD.
As a result, a single PE can be amplified sufficiently to be
safely above the noise of the FET amplifier, even when it is oper-
ating at speeds considerably higher than video rate (35–50 MHz,
vs. 13 MHz for video). As the amount of gain depends exponen-
13
In a CCD without CDS, noise features will seem to be smeared sideways, tially on the exact voltage on f2, it is possible to “dial in” the
while in one with CDS, they will appear as one-pixel-wide stipple with no amount of gain needed to keep the signal level well above the noise
directionality. of the FET amplifier. However, it is important to remember that
926 Appendix 3 • J.B. Pawley

FIGURE A3.7. Energy diagram of an electron-multiplier CCD amplifier. The high field region that occurs between f2 and fDC when f2 goes strongly posi-
tive (right) causes about 1% of the electrons passing this region to collide with a lattice electron with sufficient energy to boost it to the conduction band. Repeated
over hundreds of transfers, this process is capable of providing an average amplification of hundreds or even thousands of times.

the use of high EM gain will tend to saturate the “full-well” capac- EM-CCDs have one other important form of “dark noise”
ity of later pixels in the gain register, reducing intra-scene dynamic called Clock Induced Charge (CIC, also known as spurious noise).
range.14 Although this effect can be reduced to some extent by CIC typically consists of the single-electron events that are present
making each pixel in the gain register (and the read node) larger, in any CCD, and are generated by the vertical clocking of charge
this approach is limited by the fact that one triplet of electrodes during the sensor readout. The process involved is actually the
can control a band of silicon only ~18 mm wide and because a same impact ionization that produces multiplication in the gain
larger read-node capacitance increases the read noise of the FET register; however, levels are much lower because lower voltages
amplifier. are involved. In conventional CCDs, CIC is rarely an issue as
In sum, the gain-register CCD works like a normal fast-scan single-electron events are lost in the read noise. However, in EM-
CCD with no read noise. The high scan speed makes focusing and CCDs where the read noise is essentially zero and dark charge has
sample scanning quick and easy and the device preserves the full been eliminated through effective cooling, CIC is the remaining
spatial resolution of the CCD because the charge packet from one source of single-electron, EM-amplified noise. If left unchecked,
pixel is always handled as a discrete entity (unlike in an intensi- it can be as high as 1 event in every 7 pixels. Fortunately, it can
fied-CCD). Of course, with fast readout, there is less time to accu- be minimized by careful control of clocking voltages and by opti-
mulate much signal and the resulting image may have considerable mizing the readout process to cope with faster vertical clock speeds
Poisson noise. But this is not the camera’s fault! (down to 0.4 ms/shift). This leaves a detector with less than one
Alternatively, the output of many frames can be summed to noise pulse in every 250 pixels: a detector extremely well adapted
reduce Poisson noise or, if the signal is bright, one can turn off the for measuring zero!
EM gain and have a fully functional scientific CCD.15
If the gain-register CCD is read out fast, there is so little time
Of Course, There Is One Snag!
for dark charge to accumulate that cooling would seem unneces-
sary until one remembers that one can now “see” even one PE of The charge amplification process is not quite noise free because
dark-charge above the read noise. Because multi-pinned phasing the exact amount by which each electron in the packet is ampli-
(MPP) is less effective during the readout clockings, significant fied varies in a stochastic manner (i.e., some electrons are “more
dark charge can be generated during readout. If the exposures are equal” than others.). The statistical arguments are discussed in a
short, this source of dark charge becomes significant, and in an paper found at the URL listed below and in Chapter 4. In summary:
EM-CCD, even one electron is significant! In practice, the best as the multiplicative noise inherent in the charge multiplication
performance is obtained when the EM-CCD is cooled to between process creates noise that has a form very similar to that produced
-80° and -100°C. by Poisson statistics, the easiest way to think of its effect is to
assume that the amplifier has no noise at all but that the signal
being fed into it is half as big as it really is. In other words, the
camera will work perfectly but it will work as though it has a QE
14
If a register designed with enough pixel area to hold a normal full-well charge that is only half of what it really is. Back-illuminated sensors are
of 30,000 electrons, is used with a gain of 100¥, then the pixels near the end now available with an intrinsic QE of about ~90% or ~45% when
of the gain register will become full whenever the original charge packet has
>300 electrons. used in the gain-register mode. This is 5–10¥ better than the per-
15
Because, as noted above, the read node of the FET amplifier at the end of formance available from most PMTs especially in the red end of
the gain register in an EM-CCD has a relatively large capacitance, E2V offers the spectrum.
two separate FET readout amps. The one mounted at the end of the gain reg- It is worth noting that one can use electron multiplication and
ister is optimized for fast readout. The other is mounted at the end of the HR
not connected to the gain-register, has low input capacitance and is optimized
still maintain the full QE by using the detector in photon-counting
to read out slowly with low noise. Signal is sent to the latter by reversing the mode, as is now being done by many astronomers. Photon count-
charge transfer sequence applied to the HR. ing is only possible when one is able to confidently see a single-
 More Than You Ever Really Wanted to Know About Charge-Coupled Devices • Appendix 3 927

photon event as different from any dark event and when the on back-illuminated chips, it can reach 90% (with somewhat
number of photons collected during an exposure is so low that higher fixed-pattern noise).
there is little probability of >2 photons falling into a pixel. To The fill-factor is the fraction of the sensor surface actually sen-
implement photon-counting, one records a sequence of short expo- sitive to light. On the best frame-transfer CCDs, it can be almost
sures containing “binary”-type single-photon data, and combines 100%. On interline transfer CCDs it may be only 40%. Light not
them to generate an image that is free of multiplication noise.16 To absorbed in a sensitive area is lost, reducing the QEeff of the sensor
be useful for recording dynamic events in living cells, an extremely proportionally.
fast frame rate would be needed. This may be more possible with
Factors affecting QE:
some future EMCCD sensors.
(More info on EM-CCDs at http://www.emccd.com and Front-illuminated chips
http://www.marconitech.com/ccds/lllccd/technology.html)
• Light is scattered by the transparent, polysilicon charge-
transfer electrodes that overlie the photosensitive silicon
surface.
PART II: EVALUATING A
CHARGE-COUPLED DEVICE • This scattering is more severe at shorter wavelengths. Light
that is scattered is not detected.
• As blue light is absorbed nearer to the surface than red light,
A. Important Charge-Coupled Device Specs for and “deep electrons” may go to the wrong pixel, CCD resolu-
Live-Cell Stuff! tion may be a bit lower than the pixel count at longer wave-
lengths, especially on chips with small pixels.
Although in Part I, much time was spent discussing cooled, slow-
scan, scientific CCDs, in fact, these have not been much used in • Best QE: ~20% blue, ~35% red/green
biological microscopy since Sony introduced the ICX085, 1 k ¥ • Two efforts to improve the QE of front-illuminated chips
include “Virtual Phase” (one phase “open,” Texas Instruments,
1.3 k, micro-lens-coupled, interline-transfer chip in the late 1990s.
Houston, TX) and the use of indium oxides for the transfer
Although initially developed not for the scientific market but to
electrodes (Kodak, Rochester, NY). These have increased peak
meet the needs of the Japanese high-definition TV standard, these
QE to the range of 55%.
chips offered a set of practical advantages that biologists found
very appealing: Micro-lens array chips
— As an interline transfer chip, it needed no mechanical shutter • Sony has pioneered a process in which a micro-lens is mounted
and could be run so as to produce a continuous stream of above each pixel of a front-illuminated, interline-transfer
images. CCD. The lenses focus most of the impinging light onto a part
— The high readout speed (up to 20 Mhz) allowed real-time im- of the CCD where reflection losses are least, pushing the QE
aging compared with the 5–10 s/frame readout then common. to 65% in the green, less in the red (because of losses to the
— The 6.45 ¥ 6.45 mm pixels were small enough to sample the overflow drain) and purple.
image produced with high-NA 40¥ and 60¥ objectives.
Back-illuminated chips
— The 1 k ¥ 1.3 k raster size was both sufficient for most bio-
logical microscopy and significantly higher than that of the • Made by thinning the silicon and then turning it over so that
other scientific chips then available. the light approaches the pixel from what would have been the
— Mass production allowed the development of a micro-lens back side. This avoids scattering in the transfer electrodes and
array that increased the QEeff to an acceptable level for a front- increases the QE to about 90% in the green and >70% over the
illuminated, interline chip and did so at a price biologists visible range.
could afford. • More expensive because of the extra fabrication.
As a result, the majority of CCDs sold for use in biological • Slightly less resolution and more fixed-pattern noise, caused
by imperfections in the thinning operation, and the presence
microscopy today use this chip or its higher-QE cousin, the
of two sets of surface states.
ICX285. Although mass production made quality CCDs available
to many who formerly could not have obtained one, it is impor- Color Chips
tant to remember that the read noise of ±8–24 electrons/pixel
(depending on read speed) is substantially higher than ±2–3 elec- • One-chip color sensors employ a pattern of colored filters, one
over each pixel. Light stopped by any such filter cannot be
trons/pixel that characterized the best, slow-scan, scientific CCDs.
detected and is therefore lost. The QE of such sensors is there-
Although, as noted below, the difference is only important if the
fore at least 3¥ lower than for an otherwise comparable mono-
dimmest pixel records fewer than ~50 electrons, and this seldom
chrome chip.
occurs in widefield fluorescence microscopy, the disk-scanning
confocal fluorescence microscopes now available do provide an • 3-chip color sensors use dichroic mirrors to separate the
“white” light into three color bands, each of which is directed
image in which this difference is significant (Chapter 10).
to a separate monochrome CCD sensor. While this would seem
to ensure that “all photons were counted somewhere,” because
1. Quantum Efficiency (QE): such systems seldom employ microlenses, their effective QE
QE is the ratio of photons striking the chip to electrons kicked into is not much better than the 1-chip color sensors and alignment
the conduction band in the sensor. It should be at least 40% and of the signal light is important.17

17
While the QE is not much better, the resolution of the 3-chip camera is the
16
There is no multiplicative noise because any spike above the FET noise floor same as that of each chip, without the interpolation needed to disentangle the
counts as one electron, no matter how much it has been amplified. 3 colored images from the output of a 1-chip color sensor.
928 Appendix 3 • J.B. Pawley

• Color can be detected by making sequential exposures of a TABLE A3.2. Dynamic Range and Pixel Size
monochrome chip through colored glass or LCD filters. This
12-bit camera 14-bit camera
produces the same QE losses as the patterned filter but has w/small pixels w/large pixels
the advantage that it can be removed when higher sensitivity
is needed. This design is not suited for imaging moving Pixel Size 6.7 ¥ 6.7 mm 24 ¥ 24
objects. Full Well 27,000 345,000
Least significant bit = 6.5 electrons 21 electrons
Implied noise level ±13 electrons ±42 electrons
2. Readout Noise:
This spec is a measure of the size of the pixels and the quality of
the circuitry used for measuring the charge packet in each pixel.
It is measured in “±RMS electrons of noise” (i.e., 67% of a series on the chip determines the total specimen-to-chip magnifi-
of “dark” readings will be ± this much). cation needed!
• A good scientific CCD camera should have a noise level of Two examples:
<±5 electrons at a readout speed of 1 M pixels/second.
a. 1.4 NA 100¥ objective and a 1¥ phototube.
• The readout noise increases with the square root of the readout • The Abbe Criterion resolution @ 400 nm is about
speed (see Table A3.2).
0.22 mm. Magnified by a total magnification of 100¥,
• NO Free Lunch! A chip that has ±5 e RMS of noise when this becomes 22 mm at the CCD.
readout at 100 k pixels/sec (or 10 seconds to read out a 1024
¥ 1024 chip), should produce ±50 e RMS of noise if read out • A CCD having 8 ¥ 8 mm pixels samples such an image
at 10 M pixels/sec (or 0.1 sec to read the same chip). adequately (~2.8 pixels/resolution element).
b. 1.3 NA 40¥ objective and a 1¥ phototube.
What Is “Good Enough”? • The Abbe Criterion resolution @ 400 nm is now
0.25 mm. Magnified 40¥ this becomes, 10 mm.
Very low readout noise is only essential when viewing very dim
specimens: luminescence, or low level fluorescence. Read noise is
• A CCD having 8 ¥ 8 mm pixels is inadequate to sample
this lower-mag, high-resolution image.
only a limitation when it is more than the statistical noise on the
photon signal in the dimmest pixel (i.e., >sqrt of the number of If you must use this objective, you need either a higher mag
detected photons = sqrt # electrons). phototube (2.5¥) or a chip with 3 ¥ 3 mm pixels or (as CCD
Consider the signal levels that you plan to use. Will the darkest pixels are seldom this small), some combination.
important part of your image have zero signal or do you expect
some background signal from diffuse staining or out of focus light?
• Saturation signal level: The maximum amount of signal that
can be stored in a pixel is fixed by its area. The proportion is
If the dimmest pixel in your image represents ~100 electrons, then 600 electrons/square mm, so a 10 mm ¥ 10 mm pixel can store
the Poisson or statistical noise on this background signal will be a maximum of 60,000 electrons before they start to bleed into
±10 electrons. “Adding” an additional ±10 electrons of readout neighboring pixels. In practice, as fluorescent micrographs of
noise will not make much difference to a measurement of this living cells seldom produce signals this large, large pixels are
background signal and it will be even less significant when added usually unnecessary.
to the even greater Poisson noise present in pixels where the
stained parts of the image are recorded. However, the saturation level also represents the top end of
This is especially true because RMS noise signals add as the another spec, the dynamic range. This is usually quoted as
“sqrt of the sum of the squares” (i.e., the total noise from ±10 elec- 12-bit (4000 : 1) or 14-bit (16,000 : 1) etc., and represents the
trons of readout noise and ±10 electrons of Poisson noise is only ratio between the full-well saturation level and the readout
sqrt (100 + 100) = ±14 electrons). noise. Therefore, a camera with relatively high readout noise
On the other hand, if you are really trying to keep those cells can still look good in terms of dynamic range if it has large
alive and you find that 2,000 electrons in the bright areas is enough, pixels and hence a high full-well capacity. Conversely, a 12-bit
the dark areas may now be only 50 electrons. As the sqrt of 50 is camera with small pixels can have less actual noise-per-pixel
about ±7, an additional ±10 electrons of readout noise may no intensity measurement than a 14-bit camera with large pixels.
longer be acceptable, but only if you have to make measurements In this case, the noise level of the 14-bit camera is >3¥ that
in the dark areas on your image. In this case, the obvious choice of the 12-bit camera. Your signal/pixel would have to be
is a slower, quieter CCD or an EM-CCD. 3¥ larger in order to be “seen” when using this particular
While in widefield fluoresecence, the background stain level 14-camera.
is seldom so low that the sqrt of the signal recorded is lower than
the read noise, the disk-scanner does provide such an image 4. Array Size:19 The argument for small
(Chapter 10). As one of the main advantages of disk-scanning is
that one can scan an entire image plane very rapidly, the fact that
• Assuming 0.1 mm pixels (referred to the object plane), a 512
¥ 512 pixel chip will image an area of the specimen that is
one can read out the EM-CCD very rapidly without increasing the about 51 ¥ 51 microns. If this is enough to cover the objects
read noise makes it the ideal detector for this type of scanner (or, you need to see, this small chip has a lot of advantages over
indeed for high-speed line scanning confocal microscopes). chips that are 1024 ¥ 1024, or larger.
• Lower cost
3. Pixel Size:
• Nyquist sampling: The size of a pixel on the CCD is, in itself,
not very important BUT one must satisfy the Nyquist crite- 18
Of two times smaller than the “resolution,” as defined by Rayleigh, or Abbe.
rion: The pixels on the chip must be ~4+ smaller than the 19
The array size refers to the number of lines and pixels in the sensor, not to
smallest features focused onto it18 (see Chapter 4): Pixel size its total area.
 More Than You Ever Really Wanted to Know About Charge-Coupled Devices • Appendix 3 929

• 4¥ fewer pixels to read out, meaning either: if the signal level is so low that 1 s/frame is required to accumu-
— 4¥ slower readout clock, giving 2¥ lower readout noise. late enough signal to be worth reading out, then reducing the read
— Same clock speed and noise level but 4¥ faster frame time. time much below 0.1 s loses some of its appeal.
(Easier to scan specimen to find the interesting part! Time Faster readout speeds are particularly important for moving
is money!) specimens, especially when doing widefield/deconvolution or
• 4¥ less storage space needed to record data. when following rapid intracellular processes, such as vesicle track-
ing or ion fluxes.
The argument for big:
6. Shutter Stability:
• Manufacturing improvements are reducing readout noise Though not strictly a CCD spec, electronic (LCD) or mechanical
levels at all readout speeds, and CCDs with more pixels often shutters are often built into modern CCD cameras.21 The latter have
also have smaller pixels which can lead to lower read noise. If the disadvantages of producing vibration and having a limited life-
your labels are bright, having a larger chip allows you to see time but the advantage that they transmit all of the light when they
more cells in one image (as long as they are confluent!). are open (even an “open” LCD can absorb >50% of the light, other
Assuming that Nyquist is met in both cases, a large print of an electronic shutters may be better).
image recorded from a larger sensor always looks sharper than There seems little point in having a camera capable of record-
one from a smaller array. ing (say) 40,000 electrons/pixel with an accuracy of ±200 e (or
• Binning: Binning refers to the process of summing the charge 0.5%!) if the shutter opening time is only accurate, or even repro-
from neighboring pixels before it is read out. This increases ducible, to ±10%. If one shutters the light source instead of the
the size of each charge packet read (making it look brighter) camera, similar limitations apply.
and reduces the number of pixels. For example: 2 ¥ 2 binning
allows the owner of a 1024 ¥ 1024 chip to obtain the 7. User-friendliness:
speed/noise performance similar to the smaller chip (512 ¥ State-of-the-art cameras often seem to have been designed to make
512) and to do so in a reversible manner. However, the optical sure that no one unwilling to become a devotee of “CCD Opera-
magnification may need to be increased to preserve Nyquist tion” can possibly use them efficiently! Start off by asking to see
sampling. an image on the screen, updated and flat-fielded at the frame-scan
Before deciding that you need a larger chip, compare what you rate and showing as “white” on the display screen, a recorded
would get if the same money were spent on another intensity that is only ~5% of the full-well signal. This is where you
scope/CCD/graduate student! should do most living-cell work. Then ask the salesman to help
you to save time-series of this image. Increase the display con-
Bottom line: trast until you see the noise level of the image, both before and
after “flat-fielding.” Put a cursor on one pixel in the top frame of
• If more pixels means smaller pixels, they will each catch fewer the stack and plot its intensity over the series.
photons unless the magnification is reduced proportionally.
More pixels at the same frame20 rate mean somewhat higher 8. “The Clincher” (Well, at least sometimes . . .):
read noise because the pixel clock must go faster. Ask him/her what the intensity number stored in the computer for
some specific pixel means, in terms of the number of photons that
5. Readout Speed: were recorded at that location, while the shutter was open. To
Although readout speed has been discussed above, we haven’t answer, the salesperson will have to know the QE, the fill factor
mentioned that some good CCD cameras have variable speed read- and the conversion factor between the number of electrons in a
outs and the new EM-CCDs impose no high read speed penalty pixel and the number stored in the computer memory (sometimes
(Table A3.3). called the gain-setting). To help them out, any “real” scientific
It is convenient to be able to read out the chip faster when CCD camera has the latter number written, by hand, in the front
searching and focusing as long as one can then slow things down of the certification document (usually a number between 3 and 6).
to obtain a lower read noise in the image that is finally recorded. If the salesman doesn’t understand the importance of this funda-
However, the read speed is only one limitation on the frame rate: mental number, what hope is there for you? (Hint: It is important
because the Poisson noise is the sqrt of the number of electrons in
TABLE A3.3. CCD Specifications the well, not the sqrt of some arbitrarily proportional number
stored in your computer.)
Array size Pixel Clock Rate Noise level* Frame time Frame rate/s
640 ¥ 480 13 MHz 200 e/pixel** 0.033 30 B. Things That Are (Almost!) Irrelevant When
(video rate)
512 ¥ 512 100 kHz 5 e/pixel 2.5 sec 0.4
Choosing a Charge-Coupled Device for
1 MHz 15 e/pixel 0.25 sec 4 Live-Cell Microscopy
5 MHz 35 e/pixel 0.05 sec 20
1024 ¥ 1024 100 kHz 5 e/pixel 10 sec 0.1 1. Dynamic Range:
1 MHz 15 e/pixel 1 sec 1 This is the ratio of the “noise level” to the “full-well” (or
5 MHz 35 e/pixel 0.2 sec 5 maximum) signal. Although 16-bit may sound a lot better than 12-
bit, you need to think before you are impressed.
* Assumes conventional FET circuits. ** The readout noise is relatively higher at
video rate because the higher speed often precludes the use of various techniques, The noise level should not be more than 5 electrons/measure-
such as correlated double sampling, that reduce readout noise. ment. Period!

20 21
The readout speed of a 2 ¥ 2 binned 1024 ¥ 1024 is a bit slower than an Often the same advantage can be gained by shuttering the light source. This
actual 512 ¥ 512 because twice as many vertical clock cycles are needed, may become more common as pulsed laser or light-emitting-diode light
and one still needs to read out pixel by pixel in the horizontal direction. sources are introduced (see Chapters 5 and 6).
930 Appendix 3 • J.B. Pawley

Twelve bits is 4,000 levels. If the first level represents 5 elec- 3. “Imaging Range” “Sensitivity” (or anything
trons (in fact, it should represent half the noise or 2.5 e), then the measured in LUX):
4,000th represents 20,000 electrons or (assuming a QE of 50%), Stick to something you (and I?) understand: Photons/pixel or elec-
about 40,000 photons/pixel/measurement. trons/pixel. The other conversions are not straightforward.
How often do you expect to be able to collect this much signal
from an area of a living cell only 100 ¥ 100 nm in size? You should 4. “Neat Results”:
be able to get a good, 8-bit image using only 6% of the dynamic Unless you know how well stained the specimen is, you cannot
range of a 12-bit CCD (Fig. A3.8).22 evaluate an image of it in a quantitative manner. (Though you
As the “full-well” signal is only proportional to the area of the may not want to admit this!) By all means, view your own speci-
pixel on the chip (area in sq. mm ¥ 600), the dynamic range is only mens, but viewing “test specimens” that are not expected to fade
really impressive if it is high AND a chip has small pixels. Then and have a known structure (fluorescent beads in some stable
it means that the readout noise is low. A test for actual dynamic mounting medium?) facilitates A/B comparison. If you do use your
range is described below. own test specimens to compare cameras, be sure to view them on
Bottom line: For disk-scanning confocal microscopy, a large the same scope, and with the same conditions of pixel size and
dynamic range is only important if it reflects a low readout noise readout time etc.
level. Better still . . .
Easier to just check the readout noise!

2. High Maximum Signal (high, full-well C. A Test You Can Do Yourself!!!

number, because of large pixels): Set up each camera that you want to evaluate on a tripod, add a
On living cells, you will probably never have enough light to reach C-mount lens, and an ND 3 or ND 4 filter. Hook up a monitor or
a full-well limit of even 20,000 electrons. Even if you do, there computer and view some scene in your laboratory under ordinary
are better ways to use it (more lower-dose images to show time illumination (avoid light from windows which may vary from day
course?). to day).
Close the lens aperture down until you can no longer discern
the image (see Fig. 4.20). This is the “noise-equivalent light level”:
the signal level at which the electron signal (i.e., photons/pixel ¥
QE) just equals the total noise level. Your measure is the aperture
at which the image disappears.23 Because it is sensitive to both QE
and readout noise level, this is a very useful measure of what we
all think of as the “sensitivity.” Of course, the signal level depends
If the read noise is ±8 electrons, or 2 not only on the light intensity but also on the exposure time and
gray levels, one can obtain a
useful,“8-bit” (256 levels) image by the pixel area, so make sure to keep the former constant and make
using only 6% of its dynamic range. allowances for the latter.
If you do not have even these meager facilities (a C-mount
lens, an ND filter, a tripod and some time), take an image of
nothing. Look at “no light” for one second, and for 100 seconds.
Ask to see a short line profile that plots intensity vs. position along
a line short enough that one can see the intensity of each individ-
ual pixel. The difference in the average intensity between the short
and long exposure is a measure of the leakage.24 With a little cal-
ibration from the published full-well specs (a spec less open to
“interpretation” than “noise”), you can even get a direct measure
of the read noise level from these dark images. (It should be the
standard deviation of the values as long as they are counted in elec-
trons, not “magic computer units” and as long as fixed-pattern
noise is not a factor.) And just trying to work it all out will give
you some idea if the salesman knows anything . . .

D. Intensified Charge-Coupled Devices

FIGURE A3.8. Not using the full dynamic range of a CCD. As most scien-
tific CCDs have more dynamic range than one “needs” in live-cell fluorescence Intensified CCDs (ICCDs) are just that: the mating of an “image
microscopy, the excitation dose to the specimen can be reduced if one sets up intensifier” to a CCD. The idea is that the photon gain of the inten-
the CCD control program to display an 8-bit image using only the bottom 1,024 sifier (can be 200–2000¥) will increase the signal from even a
levels of a 12-bit image. Such an image is more than adequate for many func-
tions in live-cell biological microscopy (particularly when other factors such
as dye-loading etc., may cause larger errors) and will require only 6% as much
signal as would a “full-well” image. 23
If the lens doesn’t have a calibrated aperture ring, you can open the aperture
all the way and reach the “threshold” exposure level by reducing the expo-
sure time and adding ND filters. Remember to also correct for pixel area.
Larger pixels intercept more photons.
22 24
Remember, given optical and geometrical losses, you can collect no more With a good EM-CCD, this measurement can be done using a short
than about 3–10% of the photons produced, and, each fluoroscein molecule exposure and high EM gain, then counting the number of amplified dark
will only produce perhaps 30,000 excitations before “dying.” charge/CIC spikes across a typical line of the raster.
 More Than You Ever Really Wanted to Know About Charge-Coupled Devices • Appendix 3 931

single photoelectron above the read noise of the CCD. This occurs, • Photocathode resistivity can produce “dose-rate” effects: non-
and can be particularly useful where fast readout is needed such linearities in which the recorded intensity of the brightest areas
as when measuring ion transients. Finally, pulsing the voltage on may depend on (and affect) the brightness of nearby features.
the intensifier section makes it possible to shutter (“gate”) the
Because I expect that EM-CCDs such as those mentioned
camera on the ns time scale, making the ICCD useful for making
above will soon supplant ICCDs except where fast gating is
fluorescence lifetime measurements (Chapter 27, this volume).
needed, I have not gone into more detail here. For more info, go
However, ICCDs do not have the photometric accuracy of
to: http://www.stanfordphotonics.com/
normal CCDs for a number of reasons:
• The relationship between number stored in memory and
the number of photons detected is generally unknown and ACKNOWLEDGEMENTS
variable.
• The intensifier photocathode has low QE25 (compared to that The author would like to thank Dr. J. Janesick, formerly of the Jet
of a back-illuminated CCD). Propulsion Lab (California Institute of Technology, Pasadena,
• The “resolution” is generally only dimly related to CCD array CA), for many conversations about CCD operation and for the
size because of blooming in the intensifier. To check this, original sketches for Figures A3.1, A3.4, and A3.5 and to Colin
reduce light intensity until you can see the individual flashes Coates, (Andor Technologies, Belfast, UK) for his helpful com-
produced by single photoelectrons. See how many lines wide ments on the manuscript and for Figure A3.6.
they are. (They should be one line wide.)
• They have additional noise sources: phosphor noise, ions in
intensifier section create flashes, high multiplicative noise in REFERENCES
the intensifier section greatly decreases QEeff, etc.
Inoue and Spring, 1997, Video Microscopy, Second Edition, Plenum, New
York, 1-741, particularly Chapters 5–9.
Pawley, J.B., 1994, The sources of noise in three-dimensional microscopical data
25 sets, Three Dimensional Confocal Microscopy: Volume Investigation of
And the GaAsP photocathode with better QE, have to be cooled, making the
assembly very expensive. Biological Specimens, (J. Stevens, ed.), Academic Press. New York, 47-94.
http://www.springer.com/978-0-387-25921-5