Objectives: To understand; 1 Contrast, signal, and noise and the physical factors affecting them 2 Calculation of exposure requirements using the Rose model 3 Optimization of signal to noise ratio by proper choice of X-ray energy 4 Detective Quantum Efficiency - dependence on fractional signal amplitude 5 Minimum detectable contrast 6 Digitization noise, digitization requirements Imaging systems are characterized by: 1. The ability to detect small attenuation changes, i.e., low contrast, faint objects. This increases as signal increases and as noise decreases. 2. The ability to see small objects, fine detail. Increases as the point spread function gets narrower. 3. The ability to stop motion and to sample motion. Increases as the exposure time decreases and the sampling rate increases. This section addresses the factors which contribute to contrast resolution. These generally relate to overall system signal to noise ratio and the various factors affecting this quantity. Consider the differential attenuation due to some object as shown in Figure 1 No N1 N2 Subject contrast is defined as C = (N1 - N2) / N1 (1) assuming no scatter is detected. This may also be defined in terms of intensity. For monoenergetic beams the two definitions are equivalent. For polyenergetic beams the contrast would be somewhat different for the two definitions. In this discussion we will work with the definition in terms of fluence since we will be discussing numbers of x-rays. Christensen defines C = N2 / N1. Another definition which is sometimes used is C = (N1 - N2) / ( N1 + N2) You can use any of these but you must be consistent in the derivation of results. C =( N1 - N2)/ N approaches 0 as N1 approaches N2, and approaches 1 as N2 approaches 0. However it is defined, it refers to the difference between an object and its surroundings. When these fluences (or intensities) are detected by a detector a signal is generated, for example, video signal or optical density. Refering to Figure 2, No Object area A N1 N2 Detector resolution element of area a the input signal S observed by each detector element a is given by S = Na The size of the detector element a is determined by whatever defines the resolution of the system, for example a picture element ( pixel ) in a digital image, or the point spread function in a film image. Perception of an object in an image is dependent upon the differential signal S representing the difference between the detected signal, integrated over the object size A, and a similar area in the region surrounding the object defined by S = N1A - N2A = (N1 -N2)A {2} Whether this differential signal is perceptible depends on the degree of fluctuation on the signal in the area in the immediate vicinity of the object. These fluctuations, which may be due to a variety of causes are called noise. Noise = undesirable signal = n Noise masks small signals and, along with the differential signal, determines the contrast resolution of the system. The noise relevant to the task of object detection is the average noise over an area equal to the object size. Image quality is often described in terms of signal to noise ratio. S / n = SNR ( signal to noise ratio) The differential signal divided by the noise is also referred to as signal to noise ratio in some contexts. We will refer to this as differential signal to noise ratio. S / n = SNR ( differential SNR ) This is also sometimes referred to as contrast to noise ratio. We will not use this term since it is somewhat confusing in view of our definition of contrast as S/S. What determines differential signal and noise? Signal subject Contrast kVp filtration object scatter detector display Noise quantum noise system noise anatomical noise observer noise artifacts Consider exposure requirements for contrast and spatial resolution in an IDEAL SYSTEM - i.e. one in which all noise due only to quantum noise, ( statistical fluctuations in the number of transmitted x-rays). Xray noise is described by a Poisson distribution as shown in Figure 3. n Probability of #x-rays in area A N1 n = (N1A)1/2 = N Fluence N The differential signal is given by equation 2 as S = ( N1 - N2 ) A For very small contrasts and (3) Therefore we can write S =NCA (4) The noise is given by (5) so (6) According to equation 6, the detected fluence is related to the s/n by (7) The detected fluence is related to the incident fluence N0 by the transmission factor T as N = N0 T (8) (For non-ideal systems the detection efficiency of the detector would also multiply the transmission factor in the above equation) Therefore the incident fluence is related to the imageSNR by (9) The ROSE MODEL is a model which assumes that the statistical fluctuations calculated above are the only determinant of perception. The effects of edge gradients, which are known to be important, are not considered. The Rose model states that small contrasts will be visualized when SNR = 3 - 5 We can use the Rose model to obtain approximate estimates of required exposure. We can illustrate the Rose model using a raindrop analogy. Suppose you had a marble and a basketball on your driveway as shown in Figure 4. Figure 4 After a small amount of rain, if you removed these objects, you could detect the presence of the basketball, but not that of the marble. This is illustrated in Figure 5. Figure 5 As more rain falls, and the average spacing of the raindrops becomes smaller than the marble, you can begin to see the dry spot under the marble when it is removed. This is illustrated in Figure 6 where we have only included the region around the marble. Figure 6 A lot more rain is required to define the marble than the basketball. A higher fluence (exposure) is needed to visualize small objects. So for an ideal system, No is proportional to 1/ A, 1 / C2, 1/ T, and ( s / n )2. DEPENDENCE OF IMAGE QUALITY ON X-RAY EXPOSURE 0.025 mR Kruger and Reiderer Basic Concepts of Digital Subtraction Angiography p 80-81 0.400 mR For some time after the initial introduction of digital subtraction angiography techniques radiologists investigated the possibility of obtaining angiograms using intra-venous injections of contrast material, usually into a vein in the arm, but sometimes using a catheter placed in the superior vena cava or right atrium. Although considered less invasive than intra-arterial catheter placement, this technique required high system contrast sensitivity because of the fact that the contrast was diluted by a factor of twenty or so, depending on cardiac output, before reaching the arteries of interest. Suppose we have low iodine contrast in a vessel following intravenous injection and we want to detect 1 mm long narrowing of the vessel as shown in Figure 7. 1 mm 1 mm N2 N1 Suppose: I = 5 mgm/cm3 iodine density at the artery after iv injection t = 1 gm/cm3 tissue density t = 20 cm µtissue = µt = 0.25 cm2 / gm µiodine = µI =15 cm2/gm Keff = 35 keV Then C = (N1 - N2) / N1 = (e-µt * 20 - µ t µt*20∆ I I) e- / (e-µt* 20) = 1 - e - µI ∆tI ~ 1 - ( 1 - µI ∆tI .... ) Therefore, in general, for small contrasts (10) Plugging in, we get C = µI ∆tI = µI pl ∆xI = 15 cm2/gm* .005 gm/cm3* .1 cm = .0075 The transmission factor is Plugging into equation 9 with s/n = 3 we get Remembering that 1R ≈ 2x1010 photons/cm2 at 35 keV we get a required incident exposure of E0 ~ .12 R This is the exposure required to image the 1mm length of a contrast filled vessel. This should also be approximately the exposure required to see a total stenosis with a length approximately equal to the vessel diameter. This shows that exposure is not only related to contrast resolution, but also can limit spatial resolution, i.e., in order to see something very small (of low contrast), a large exposure is required. For example, what if ∆tI decreases to 0.1 mm, down by a factor of ten? Then contrast decreases by a factor of ten to 0.0075, area goes down by ten, assuming we are still looking at a 1mm length, and T stays the same since it is dominated by the tissue transmission. Then N0 = 2.39 x 109 x 102 x 10 = 2.39 x 1012 E0 = 119 R ! The effect of increasing the number of quanta in an image are illustrated in Figure 8 for the case of light photons. This is from Rose’s book called Vision. 3 • 10 3 ROSE W0MAN 9.3 • 10 3.6 • 10 4 5 1.2 • 10 4 7.6 • 10 5 7 2.8 • 10 An illustration of the same principle using x-rays is illustrated in Figure 9. Figure 9A shows an attempt to image the arteries in the head using an intravenous injection. For this image there was a malfunction of the exposure regulation circuitry and the exposure was approximately a factor of 100 less than anticipated. Figure 9B shows the examination repeated with the correct exposure. The increase in image quality is evident. A Figure 9 B We can use the Rose model to calculate optimal beam energies for various situations. Again, we assume quantum statistical noise dominates image noise. We will try to determine the optimum x-ray energy for a particular imaging problem. CASE I: Small tissue thickness t on a larger tissue thickness t (gm/cm2) as shown in Figure 10. N0 t ∆t N1 N2 From equation 9 No ≈ (s /n )2 / C2 AT so for fixed s /n and A where µt is the effective attenuation coefficient over distance t, ignoring the details of beam hardening for now. but Therefore, Now we can find the beam energy which minimizes exposure which is No using the usual calculus max / min procedure of setting the first derivative to zero. No (e µtt) / µt 2 (dNo) / (dµt) = (teµt t) / (µt 2) + (-2eµt t) / (µt 3) = 0 (11) (12) which gives µt = 2 / t That this is a minimum and not a maximum can be checked by looking at the second derivative. (13) Chest imaging: t ≈ 20 cm The attenuation coefficient associated with the optimal energy is given by the above equation as µT = 2/20 = 0.1 cm2/gm This corresponds to an effective beam energy of about 100 keV and a kVp of something like 200 depending on the filtration. Mammography t ≈ 5 cm The same analysis results in µt = 2/5 cm2/gm Keff = 25 keV for tissue and 23 keV for fat Bear in mind that these are ideal system results. For actual mammography, where object contrast must be kept somewhat higher than the ideal to compete with addition system noise such as film grain, the optimal energies actually used are somewhat lower. Daffodil Imaged at 5 and 20kV 5 Gilardoni et al. Radiology-Electromedicine 20 From the examples, we see that the best energy is very thickness dependent. There is a tradeoff between transmission and contrast. For large tissue thicknesses, transmission is the most important factor, i.e., it is desirable to use the highest kVp possible until the contrast is so small that it becomes comparable to the system noise. In computed tomography, with high device SNR, studies show 140 kVp to be superior to 80-120 kVp for tissue tumor detection. Refer to Figure 10 below, with a small iodine thicknesstI substituted for the small tissue thickness t. In this case, for fixed A and tI, the required fluence is given by No ≈ (s/n)2 / (C2 1 / (C2 e-µt t) N0 t ∆t N1 N2 (14) but since C ≈ µItI No (e+µt t) / (µI2∆tI 2) (e µt t) / µI 2 (15) Now we have an expression involving two attenuation coefficients, both of which are dependent upon energy. Therefore, the differentiation involved in the minimization procedure must be with respect to energy. In order to accomplish this we will have to parameterize the attenuation coefficients in terms of the energy k and fill in the appropriate terms in the expression. dN / dk = (N / µI) (dµI/dk) + N /µT)(dµT/dk) = 0 (16) For tissue the Compton and photoelectric contributions are both ≈ 0.25 cm2/gm at 25 keV. Modeling the Compton contribution as a constant with energy and inserting the known energy dependence of the photoelectric component we get µT ≈ 0.25 [1 + (25/k)3] (17) For iodine we can make the approximation that the only significant contribution is photoelectric because of the high atomic number of iodine. Normalizing to the value of 36cm2/gm at 33 keV we get, µI ≈ 36 (33/k)3 valid for keV > 33 (18) From these expressions for the attenuation coefficients we get (dµT / dk) = (0.25)(25)3 (-3)/(k4) (dµI)/(dk) = (36)(33)3 (-3) / k4 Inserting into equation 16, we get, dN / dk =[ e µt t (-2) / (µ3I) ](36)(33)3 (-3)/(k4) + [te µt t / µ2I] (0.25)(25)3 (-3) / (k4) = 0 Simplifying, [2(36)(33)3] / (µI) = t (0.25)(25)3 Substituting for µI gives, [2(36)(33)3] / [(36)(33)3/k3] = t (0.25)(25)3 solving for k we get, k3 = t [(0.25) / 2] (25)3 k = 25 (0.25/2)1/3 t1/3 k = 12.5 t1/3 for k > 33 keV (19) For k < 33 keV the model does not apply. At 33 keV t = 18 cm. For t < 18 cm 33 kev will minimize the exposure. For t > 18cm the above formula predicts that the optimal energy increases away from the k-edge very slowly with thickness. For example at t = 27 cm, k =12.5(27)1/3 = 12.5(3) = 37.5 keV This dependence is quite different from that of tissue where the optimal energy for detecting small tissue contrasts was considerably higher at the same tissue thickness, e.g. CASE I (tissue detection): t = 20 k ≈ 100 keV CASE II( iodine detection): k ≈ 34 keV In our discussion of ideal systems we only considered quantum noise. In real systems additional noise can be contributed by the imaging or display components such as television cameras or film. Let us assume that the imaging system transforms an input fluence to an output signal as illustrated in Figure 11. Input x-ray signal N + quantum noise Imaging system Output signal S + quantum noise + system noise It is useful to define a quantity called Detective Quantum Efficiency , DQE, which is a measure of how well the system utilizes the incident x-rays to produce an image. This concept includes the actual detection efficiency of the detector which we will designate as . The DQE is defined in terms of the input and output signal (S) to noise (n) ratios S/n as DQE = (S/n)2out / (S/n)2in (20) Note that in the above equation S represents the entire signal and not the differential signal associated with any particular object. The input signal is given by , the input fluence multiplied by the characteristic area a of the system image element e.g. pixel (picture element) size. Sin = = Na (21) The DQE we calculate will depend on the size of a. Choosing larger a will increase the averaging of system noise and increase DQE. In some analyses, DQE is actually plotted as a function of spatial frequency, which will be discussed later. For the purposes of this discussion we will just assume a fixed fundamental image element size a. The input quantum noise is given by the square root of the number of x-rays in a. nin = 1/2 22) giving (S/n)in = 1/2 (23) Therefore, DQE = (S/n)2out / = NEQ / where NEQ is called the Noise Equivalent Quanta and represents the number of photons an ideal system ( with DQE =1) would have to detect to produce the same output S/n as the real system. Because the real system detects only a fraction of the incident x-rays and contains system noise, NEQ is always less than the actual number of incident xrays. In order to illustrate the role of detection efficiency and system noise on DQE and the detection of objects in an image, we will now consider the details of a video camera detector such as those used for fluoroscopy or digital radiography and digital angiography. In these applications, the transmitted image is detected by an image intensifier, the output of which is viewed by a television camera tube. The output video signal is shown in Figure 12. noise Vm blanking Sync pulse Superimposed throughout the video signal is noise (illustrated only at the top of diagram). The total noise is due to a combination of the camera noise Vc and the electrical representation of the quantum noise. We will look in detail at how these are combined a little later. The dynamic range DR of the camera tube is defined as the ratio of the peak signal Vm and the camera noise. DR = Vm / Vc (25) Modern cameras typically have dynamic range values of 1000 or more. Cameras typically have a linear response over a certain range of exposures. In order to reduce the role of the video camera noise, the light incident upon the camera tube is adjusted by means of an aperture placed between the image intensifier output and the TV camera input so that Vm is near the top of the linear range . aperture patient Vm video signal image intensifier tv camera large aperture (low dose/image) small aperture (high dose/image) detected X-rays/pixel The relationship between camera signal and, the number of detected photons per pixel (in this case determined by the bandwidth of the electronics) is shown below. The maximum signal Vm corresponds to the maximum number of photons per pixel m in the linear range as shown in Figure 13. Saturation region Vm Camera signal noise Detected x-rays per pixel m Figure 13 The output voltage (not including noise) is related to the number of detected x-rays by V = CC 0 (26) where is the detection efficiency. The error in V due to quantum noise is given by q=C 1/2 (27) The x-ray to voltage conversion factor does not get included in the square root because it is just a scaling factor. Intuitively, it might be expected that q should be times the noise in the incident x-rays rather than the square root of the number of detected photons. Professor VanLysel will discuss the reasons for equation 27 in the CT/DSA course The error in V due to camera noise is given by C = Vc = Vm/DR (28) The overall noise variance is the quadrature sum of these. n2 = 2total = 2q + 2C = C2 + Vm2 / DR2 From equation 26 we have C = Vm / m so, (29) (30) The output signal to noise ratio is then given by equations 26,29 and 30 as, (31) Now we can calculate DQE. The input signal is given by Sin = / The input quantum noise is given by (q)in = (/)1/2 giving (33) (S/n)2in =/ From the definition of DQE (eq.24) we obtain, (34) The m / DR2 term is a constant for a given aperture. In other words for a given TV camera and aperture,m is determined. Therefore the aperture is a means of adjusting the exposure and therefore the amount of noise in the image. For large apertures, the maximum signal is achieved with a small x-ray exposure and quantum noise will dominate over camera noise. (The percent noise is largest at low exposure). For small apertures, a large exposure is required to achieve maximal video signal and the fractional quantum noise fluctuations will be small. In this situation, camera noise becomes a larger factor. Lets consider the minimal detectable contrast and DQE for large, intermediate, and small video signals. V= Vm, ( = m) Assume DR = 1000, A = 1mm2, and that Edetected = 1 mR produces Vm. Then m = 2 x 107 x-rays/cm2 x 0.01 cm2 = 2 x 105 x-rays per pixel. Plugging into equation 31 we get Therefore n/S = .0025, or n = .0025 S The Rose criterion for signal detection says that the differential signal S must be 3-5 times larger than n. Taking S = 4n, the minimal detectable contrast is given by Cmin = S/S = 0.01 i.e. , a 1% contrast can be seen at a 1mR detected exposure for a 1mm2 object in the brightest portion of the image. From equation 34, assuming that the x-ray detection efficiency = 0.5, we get V= Vm /10 In this case giving n/S = 0.012 and, according to the Rose criterion,S=4n ~ .05S and Cmin = S/S = 0.05 Therefore 5% contrast is required for detection at V = Vm / 10. The DQE in this case is reduced to 17%. V = Vm / 100 Using the same equations, it is found that the minimum required contrast is increased to 40% and the DQE is reduced to 2.4%. This points out the importance of removing bright spots by means of bolus materials such as water bags or other devices. Figure 14 illustrates the effect of bright spot removal. Without bright spot removal, the video aperture or the exposure must be adjusted so that the central bright spot must be at the top of the unsaturated video range. This relegates the signal corresponding to the artery, which is superimposed on a thick region, to the small video signal regime where SNR and DQE are poor. When the bolus material is added the transmitted image becomes more uniform and can be brought up to the maximum video level either by means of an exposure increase or enlargement of the video aperture. In either case, the larger arterial signal will compete more favorably with the camera noise. Without brightspot removal With brightspot removal “artery” “patient” video “patient” + bolus Amplified video The bolusing operation in Figure 14 alters the large area contrast of the image by removing the contrast of the area creating the brightspot. However, in applications such as blood vessel imaging (angiography), especially when subtraction techniques are used to remove nonvascular anatomy, the net effect is improved vessel signal. As we will discuss later, the scatter fraction measured at the vessel will also be reduced due to the reduction in scattered radiation from the highly transmissive region. Figure 15 provides another illustration of the role of bright spots in setting up the peak video signal via aperture or mAs changes. In Figure 15A there were lines of high intensity coming through cracks in the phantom which consisted of a barium strip placed,over a pressed board step phantom. When the peak video signal was set using this large signal, the rest of the image information was placed in the low signal amplitude portion of the image and noise became dominant as shown in Figure 15A. In Figure 15B the bright spots were simply ignored and the peak video was set to match the rest of the image. This resulted in improved signal to noise ratio for the barium strip. Effect of Aperture Bright Spots at Peak Video Bright Spots Ignored A further illustration of the effects of system noise is shown in Figure 16 where in 16A a video camera has been operated with the brightest portion of the video at peak video signal and in 16B the brightest portion of the scene has been placed at 15% of peak amplitude. This simulates the difference in image quality that would be perceived using cameras differing in dynamic range by a factor of 6.6. 100% 15% Figure 16 A B As x-ray exposure is increased, fractional quantum noise is reduced relative to camera noise and image quality is improved. It is important to appreciate that as exposure is increased the quantum noise, which is the square root of the detected exposure, increases in absolute value. However, when the optical aperture is reduced to keep Vm fixed, the percentage of the video signal occupied by quantum noise is reduced. Since the camera noise is not affected by the aperture, it becomes more dominant at increased exposure and further increases in exposure produce smaller gains. Consider the behavior of the DQE at maximum video signal. Consider the case of V = Vm. From equation 31 for the signal to noise ratio we get Figure 17 shows the SNR at Vm for two different camera dynamic ranges. The 1000:1 camera has low noise and the 250:1 camera has high noise. In the case of DR = 250 where system noise becomes dominant very quickly as exposure is increased, it makes little sense to increase exposure beyond 0.5 mR per video image. For DR =1000, image quality continues to improve with exposure, leveling off at much higher exposures. Figure 17 DR1000 1000 DR DR = 250 For modern image intensifier-television based digital angiography systems such as used for digital subtraction angiography (DSA), detected exposures on the order of 0.5 mR are common. For cardiac cine’ angiography where 30-60 frames per second are typically recorded on film or real time digital recording devices, the exposure per frame is typically .05 mR per frame and the images are dominated by quantum noise. The exposure dependence of the DQE is illustrated in Figure 18 for V =Vm. In this case DQE is given by Figure 18 DR = 1000 DR = 250 Notice that for the DR =250 camera, DQE falls off quickly with exposure signifying decreased gains in image quality as exposure is increased. Because of the large camera noise component additional x-rays are less inefficiently utilized than in the DR = 1000 camera. For the previous decision we have looked at the maximum video signal. The lower signal portions of the image will already have lower DQE at any given exposure and will improve SNR less rapidly than the maximum signal region as exposure is increased. However, it is usual to try to set the exposure based on the large signal region of the video signal and try to use bolus materials to make the video signal distribution as uniform as possible. The quantum sink of an imaging system is defined as the point in the system where the signal corresponding to quantum fluctuations reach their largest fraction of the total signal. Ideally, the quantum sink should be at the front end of the system where the x-rays are detected. In a well-designed system, the relative importance of quantum fluctuations due to x-rays or, for example, optical photons produced by the x-rays, should not become more dominant in a fractional sense than they are at the point of x-ray detection. Let’s look at two examples to illustrate this concept. Image-intensified fluoroscopy Figure 19 The number of detected x-rays for an area A is given by ( = NA) (35) Following detection, the x-rays produce large numbers of electrons. The number of electrons per x-ray is sufficiently large that the quantum fluctuations are completely dominated by the x-ray fluctuations. When these electrons impinge on the output screen of the image intensifier, light photons are produced. The number of these associated with each incident x-ray is still sufficiently large that the fluctuations in the number of light photons subsequently presented to the video camera are dominated by the initial fluctuations in the x-rays detected at the input screen. The combination of electron production and subsequent photon production leads to an overall gain factor G, giving a number of photons associated with input screen area A of (36) The signal to noise in and the signal to noise out are both equal to Because of the large numbers of electrons and photons produced, the quantum sink for the case of image intensified fluoroscopy remains at the input of the system as it should. Non-Intensified Fluoroscopy Before the introduction of image intensified fluoroscopy as shown in Figure 20, radiologists looked directly at the output of a fluorescent screen detector. This had several disadvantages. Because of the low light levels, dark adaption was required and rod vision, with its lower visual acuity, had to be employed. Furthermore, as we will see, the quantum sink moved from the input screen to the eye of the observer, indicating that the information perceived was less than that available at the fluorescent screen. Figure 20 0 [ 0 ± √ (0)]• g R screen Eye The factor g corresponds to the number of light photons per detected x-ray. Since this is large, the quantum fluctuations after the screen are dominated by the x-ray quantum statistics as in the case of the image intensifier. However, following the conversion to light, which is spatially isotropic, the eye, which we will assume has area Ae, intercepts only a fraction of the total light photons. This fraction is equal to the ratio of Ae to the area of the sphere of radius R. Therefore, the number of photonse detected at the eye is given by (37) This number can be less than one per x-ray depending upon the brightness of the screen.The signal to noise ratio is (if # photons/xray is still large) at best (38) which is the SNR after the detector multiplied by the second square root factor . Therefore the quantum sink moves from the screen to the eye whenever (39) (in which case S/n is worse than eq. 38) So far we have discussed two general sources of noise, quantum noise and system noise. In many systems the analog signal produced by the system is digitized. This is the case in DSA, phosphor plate imaging, computed tomography (CT) and many other applications. Since the digitization process also introduces a sort of noise, it will be useful to break the system noise into its analog component a, including quantum noise and analog system noise, and the digital component d. Digitization is accomplished by an analog to digital (A/D) converter, which transforms the analog signal into 2N levels where N is the number of digital bits in the output of the A/D converter. For example, for an eight bit converter the analog signal is converted to 256 discrete levels from 0 to 255 as shown in Figure 21 for the case of a single horizontal line of video which has been chosen to have a monotonically increasing (ramp) signal for the purposes of illustration. The analog video signal Va is on the left. The digitized signal Vd is on the right. Figure 21 Analog Digital The digital spacing, which we will designate as , must be chosen on the basis of the amount of analog noise in the system as we will discuss below. The assignment of a given analog voltage to a digital level is made on the basis of voltage thresholds. As illustrated in Figure 22, if the analog voltage is equal is greater than or equal to /2 above the previous level, it is assigned to the next digital level. Because of this the digital representation of the image is discrete and in some circumstances can have a blotchy appearance in which large areas of the scene get caught below a threshold and then, at some adjacent point in space, abruptly make the transition to the next level. Figure 22 i+1 u Analog signal i i-1 Digital levels Thresholds We can calculate the RMS error associated with the digitization process by considering the error associated with each digital level, indicated by u. If we assume that the analog signal has equal probability of falling between two digital levels, then we can calculate the RMS digital error dig by integrating over u as follows. The total noise variance including analog and digital noise is given by the usual quadrature sum (41) The ratio of the total noise to the analog noise as a function of /a is shown in Figure 23. For small digital level spacing, the digitization process adds little noise to the analog noise. As the level spacing increases, the digital noise can greatly increase the total noise. Usually the criterion that (42) is chosen to ensure that the digitization process increases overall system noise by less than 20%. The number of bits needed to satisfy the above criterion depends on the system analog noise and the range of signals to be digitized. Referring to our video camera example, suppose that we wish to digitize the video signal between the levels Vm and Vm/10. It is common to choose such a lower level below which one might not hope to see much anyway because of the dominance of noise in the lower signal regions. For the case discussed earlier with a detected exposure of 1 mR , a 1mm2 pixel area and Vm/10, we had a = .012 ( Vm / 10 ) If we use the criterion that = 2a= .0024 Vm Then the number of digital levels #d required is given by Since 28 = 256 and 29 = 512, this requires a 9 bit A/D converter. Notice that the level spacing criterion must be applied where a is the smallest, in this case at Vm / 10. If the criterion is satisfied in the small signal portions of the image, the criterion will be satisfied in the large signal regions where the absolute size of the noise a at Vm is larger because of the larger absolute size of the quantum noise contribution. These concepts are illustrated in Figure 24, which shows a series of cardiac images digitized with increasing numbers of bits. For the lower numbers of bits, the signal in the dark portion of the image appears blotchy due to the fact that the level spacing is too large. As the number of bits is increased, this blotchiness disappears. Notice that in the bright regions of the image there is no blotchiness because the absolute size of the analog noise in these regions is large enough to satisfy the level selection criterion. Basically the noise helps signals jump between levels so that large areas do not get stuck in a single level over a large spatial region. Figure 24 ADDED NOISE NONE 0.8 % 3.1 % 12.5 % 50 % BITS 8 7 5 3 1 Figure 25 illustrates the effects of having an insufficient number of digital levels for the case of a digital chest film.