 What is now the sensitivity of a detector? We are looking at all these parameters. So, we define sensitivity as the charge that is produced for incident x-ray quantum at a specified energy because the sensitivity can be different at different energies. And therefore, I mean if you have detectors that are sensitive to different energies, you could actually overlap them. So, what is the dynamic range? So, that again is a ratio where x max is the x-ray fluence producing the maximum signal. x noise is the fluence producing a signal equivalent to the quadrature sum of the detector noise and the x-ray quantum noise. Basically, the dynamic range depends on the smallest thing that you can detect which is typically the noise. And what we do is we try to equate the RMS noise of the signal to the RMS noise of the least significant bit of the A to D converter then you have an optimum system. Let us talk about MTF. So, essentially the concept of MTF actually is a description of the resolution properties of an imaging system described in terms of the ability to produce images of sinusoidally varying test objects of various frequencies. These are sinusoidally varying objects and many times we do it with different slits that have distances or we even do it with edges because it is in the transform domain. So, essentially what we have is n sub x and delta n sin kx where k is the angular spatial frequency which is related to the frequency with the term 2 pi. It is related to the spatial frequency. The contrast associated with sinusoidal waveform as we define it is delta n which is the height of the sinusoid above the mean divided by the mean that is what we just. So, we have things called the contrast transfer functions as well. We have n sub n going into an imaging system that gives us n sub out. So, you have n sub out is a constant multiplied by n bar which is the dc part of it and the MTF as a function of k multiplied by delta n multiplied by sin k of x and the phase. So, the factor MTF k modifying the magnitude of the output signal we call it the MTF and it is the ratio of the output to input contrast as a function of the spatial frequency k. And obviously, there is a difference between linear and angular frequency and everybody I am sure knows about it. So, physical objects can be represented as a weighted sum of spatial frequency components. It is actually a 2D Fourier transform representation could also be a wavelet transform people do that that depends on the transform space that you choose to work in. See, because M sub k varies with k obviously, you will generate a distorted version of the actual object because some of the high frequencies may be attenuated. And so, therefore, you need to have a system MTF that is comparable with the task at hand or the exam that you are doing. Now, for example, if you are looking at lung nodules, you need to make sure that you have a few line pairs per millimeter in order to visualize the nodules. And then on the other hand, if you are looking at micro calcifications, you need a larger MTF. So, essentially this is another view of MTF related to the optical transfer function. Basically, if you look at it, if you have Tu, v normalized by T00 which is the optical transmission transmitted by the system gives you another view of the MTF. This is from optics for people who like to relate it to optics. An ideal device, if we have is a perfect photon counter. So, your signal to noise ratio is the number of photons divided by the square root of the number of photons which is the noise. So, that is N mean square root of N mean. So, N measured is signal to noise ratio square of the measure. So, typically, noise is not really Poisson distributed you know. And detectors, unless they are photon counting detectors or photon integrated detectors, so it gets modified. So, it is less than the signal to noise ratio. So, we have a term called N prime which is called the noise equivalent counter. And this was a tough thing initially when I was exposed to it about 20 years ago at Kodak for somebody who is used to electrical engineering terms. This was counterintuitive, you know, this is almost like 1 over the signal to noise ratio. So, number of quanta that would produce the signal to noise ratio non-ideal in an ideal system. So, if it is less quanta that produce the signal to noise ratio, it is less. But you know a higher signal to noise ratio means less noise equal equivalent quanta. So, it can be a little confusing. And so, the ratio of the noise equivalent quanta and the actual number of incident photons is called the detective quantum efficiency. It is called the DQE. It is the fraction of photons contributing to a measure result. To give you an example, in a screen form system you get a QE of 35 to 40 percent that means only 35 to 40 percent of the X-ray photons that hit the screen form system are converted to light photons. And the QE of film is even quite low. It is about 4 to 5 percent actually film does not have very high QE. So, a little bit about the electronics I mean I will try to make it as intuitive as possible. So, essentially what happens is you have this is the representation of the photodiode. This is an equivalent circuit and you have this is the pixel capacitor, the TFT switch and you could have a bunch of these things. I showed 2000 of them. I mean there should be 2000 of them. I could not show 2000 of them naturally. Then you have things in terms of cross talk from other columns, you have the column capacitance. And then we have a thing called a coordinated double sampler that I will talk about which actually came from charge couple devices, you know that technology. They have a multiplexer and anti-aliasing filter and an AD converter. So, let us discuss these components very quickly. So, each of these components has attributes which are the dynamic range. The noise actually it is interesting. Electrical engineers like to talk in terms of walls and RMS noise and physicists like allotrons. In fact, it took me a while to be comfortable with allotrons. Now, I am comfortable with both because I have to deal with a lot of physicists and they never understood what I was talking about and I had to basically translate noise. And as we know the noise is bandwidth dependent. So, in the next few slides we will look at a 2000 by 2000 array with a 14 bit conversion and see how the signal and noise propagate. These numbers are examples. These are not exact numbers, but they are representative of what you see in a digital radiography chain. Let us look at noise in a time invariant system. So, if a signal with spectrum S sub S F is applied to a LTI system with a transfer function of S sub X, then the output spectrum obviously is given by its you know S Y F is S X F times the square magnitude square of the transfer function as a functional frequency. This is a well known relationship. So, let us look at photodiode in the storage capacitor. Let us take a look at the bias voltage of 5 volts. Typically, what happens is when allotrons this is I have shown this thing as if the charge is positive, but usually what happens is the bias is extinguished in the diode. So, it may be negative charge, but for the purpose of explanation it does not matter. So, let us say the initial noise is 500 allotrons or 80 micro volts RMS. So, let us say the signal level is 4 volts. This is a fairly intense system for a dynamic range of 50,000 to 1. So, let us say the on resistance of the TFT is about 3 mega ohms and I took off to on ratio of 10 raise to minus 5 or 10 raise to minus 6. Actually, these days we try to get to 10 raise to minus 7 and 10 raise to minus 8 and the leakage is about 10 femto amps because the leakage is also important because it charges the pixel capacitor and then you have you know the control input which is VCN and then you have drain to gate capacitance and gate to source capacitance. These things actually inject charge when you move the control inputs. So, charge injection due to these capacitance this can be about 2 tenths of a volt. So, essentially you need high dynamic range so that you can subtract out these constant factors because essentially the signal can be about 12 and a half to 30 and a half bits. So, there is a thing called KTC noise and it can be proven and it is there in textbooks of electronics. So, when a switch is opened there is always noise accumulated that is called Q square is KTC where K is the Boltzmann constant T is the Kelvin temperature and C is the charge. So, C is the capacitance. So, essentially if you visualize it voltage square is KT over C and this noise is independent of the resistance of the switch it is a simple proof it can be found in any books on communication theory or microelectronics. So, column capacitance can be 50 to 100 picofarads the larger the longer column is the larger the capacitance and so, crosstalk capacitance between columns can be a few hundred femto farads. So, I represented as this. So, now here is the thing where does this column capacitance come into the picture? You got this 0.5 picofarads here and then you have 50 picofarads these are two impedances into an op amp. So, the noise gain of this op amp is 1 plus Zf over Z sub i and so, essentially the noise gain for the op amp is related to the column capacitance. So, any noise gain here is going to be amplified. So, it is one of the biggest sources of electrolyic noise in the system and you need about 10 time constants or 3 microseconds to completely charge an integrating capacitor to a 15 bit accuracy. So, the noise is integrated in 3 microseconds and it is about 1000 electrons and each of these amplifiers deals with 128 columns for an available ASIC. So, I took a very pristine amplifier that has got 2 nano volts per square root hertz but actually in practice it is not that low because we cannot run these amplifiers at very high currents it is more like 14 to 20 nano volts per root hertz. So, you see the effect of the column noise is even more severe. So, essentially in terms of the operation of the system you reset it the pixel capacitance is reset to a reference voltage and then you integrate it the signal charge is accumulated on the pixel capacitance. Then in the readout the signal charge is transferred to an external charge amplifier and then you convert it. If you look at what happens is if you start counting all these electrons here you get 1500 electrons of random noise that is 800 electrons of KTC and then 256 25000 electrons of signal and then you got an offset. So, you have total voltage of 8 volts signal signal 0.4 volts offset and 66 nano volts due to KTC and then you also have 1 over f noise of a mob of silicon transistors that are quite high. So, in correlated double sampling what you do is a sample is taken with the TFT Q1 off and immediately another is taken with the TFT on. So, the 1 over f noise is removed because you do it fairly quickly and then you also eliminate the noise due to the reset of the amplifier. But what happens is because you have an extra amplifier your square root of 2 times the RMS noise, but it turns out that this is still an advantage. You have total voltage is 8 volts 0.4 volts of offset 32 nano volts due to KTC 80 nano volts square squared noise and no 1 over f noise. So, we use these and these are actually within the chip itself each one of them as a correlated double sampler and then let us assume 14 bits with a 10 volt range. And so, the quantization noise of a bit is sigma square over 12 that is 31 nano volts. And so, if you normalize the gain to 1.2 basically the total RMS noise is the square root of the expression in the parenthesis or 398 micro volts. This is less than 610 micro volts which is the step size of the A to D this is an ideal case in a true situation that is not the case. You know, but then you are taking away some of the dynamic range like 800 counts of this 14 bit converter go away in terms of offsets and I am not address crosstalks here that can be done. So, these are the timing assumptions. So, you have a total time of converting a pixel as 3.6 microseconds. So, if you take 2000 pixels per column and 128 channel ASIC it is 920 milliseconds if you have a single A to D converter for that. So, the only way to improve this timing is to have multiple charge amplifiers per ASIC or have multiple converters and that is done and that becomes expensive. So, typically to give you an idea the raw cost of a panel may be about $10,000 the cost of the electronics is another 4 and then there are markups. So, high performance charge amplifiers are necessary to maintain signal to noise ratios. The on resistance of the TFT has to be reduced, but in order to reduce the on resistance of the TFT you need larger length to widths or larger widths and then if you do that you are taking away the fill factor. But these days we have 3 dimensional structures, but you are also increasing the leakage current. So, it is actually a balance of all these factors and these are the things that I struggle with when I design a system as to what the right balance is and that is where a lot of my time is spent and that is the reason for my gray hair. Let us look at this. This is actually some examples courtesy of robots research center in London, Ontario, Canada where you extract organs from a 3D abdominal CT image. So, if you look at it one is the vertebrae two are the ribs I have numbered all these not going to read these you can read these and this is a volume view. So, how do you do this? You do bone segmentation and actually this is in the purview of segmentation you extract the bones you separate the spine from the ribs then you find the 12th rib which is a reference then you find the vertebra connected to the 12 rib and then you set the center of gravity of the vertebra as the reference point which is the origin of the coordinate system. Now what happens is then you have a pattern spectrum and you do a size analysis and if you look over here basically you find them clustered very nicely stomach of course is misspelled I did not do this slide. What you do is you have a before and after you actually you do some processing here which I want go into details and what I want to show is if you have four regions here that you are cutting regions you show these cutting regions in 3D a lot of it is also used for image guided surgery. And then you have a histogram then you start thresholding these things applying functional transforms and then you get a result here and this is highly sophisticated it is a threshold segmentation but you have to find the right threshold because you see you see two peaks here in the histogram sometimes it is not that cut and dries easier to do this on cadavers than real people. And so the same thing I mean you use for the people who are into image processing use things like recursive erosion geodesic influence and region syncing to get all the seeds and then you recover the separated organs. Actually there are a lot of handbooks on this I can recommend you to some of these things for those who are interested after this is done but I am just giving a smattering and an example of what is being done. And then you have an organ recognition but what do you have to do you do a feature analysis in terms of size and position you match the database or dictionary because there is no such thing as a standard organ you have an atlas of different organs that actually is also culturally different you know you have atlases for organs from people from Asia or Japan also the European organs they are they are very different if you do not have these atlases you can get false recognition and you label the objects with a unique symbol a number or an anatomic name. So that is an example of that now we talked about display of images okay coming back to the display of images today all images are typically they are not displayed on film or displayed on LCD monitors they used to be displayed on CRT monitors but there is not a single CRT monitor that is used in medical imaging. Today you have medical displays that go anywhere from 5 megapixels to 9 megapixels and one of the big things before I talk about those things let us talk about what is the what are the factors that you need for such displays you need to have sufficient spatial resolution because it is not good enough to have good spatial resolution in a detector in order for that image to be aliased by a display system because the display system can alias it as well you need sufficient dynamic range which is bit depth you know you need to the ability to view pathologies and by the way early displays had a problem because they did not have sufficient dynamic range and when people did studies of radiologists a typical radiologist was able to read a normal screen film in 28 seconds and say next now they can do that that fast today with other transformations but in the early days of windowing and leveling about 10 years ago it took about two minutes to do the same thing on a display and time is money and so obviously you couldn't do that and productivity is very important and you need sufficient sharpness and frequency response and you don't want a displayed image to create a pathology I know I put all these things in order to inject some humor but this is a little more than humorous when it comes to some of the situations in the states where basically even companies can be sued for example as a five manager project in a company and let's say I do something that causes injury or death I can be hauled in the court along with the CEO of my company and the display should be bright enough and large enough to view against normal lighting condition what I mean normal lighting is not this lighting this is normal lighting in a radiologist suite because there's a little bit of darkness but it's not so dark that it puts people to sleep and it causes so much of eye fatigue and the other thing that displays to is the angle of viewing displays been a lot of studies in medical conferences where at a 45 degree angle the display looks terrible and today they're displays that use that are coming up with organic LEDs that's again a Kodak invention where the LEDs are right on top and it's isotropic with angle the advantages a bunch of people can view the same case and see the same image in order to get opinions so let's go with a few concepts but this is a highly simplified thing because if you want to look at the contrast sensitivity of the human eye the seminal book on that is by Barton's called contrast sensitivity of the human eye I think it was written in the early 2000 by Barton's or it's Barton B or T and then you can google it and find it where the contrast sensitivity has been studied in terms of various factors but this is an early study of contrast sensitivity in the luminance space is you know if you have an object and it's around the contours between the rectangles will not be visible if you have delta L over L less than 0.012 that translates in the density space optical density as 0.005d and why is this important let's say you digitize something as we will see with incorrect digitization you can actually create contours and what happens is if these contours are obvious you know people can discern the difference between what you say anatomy and a contour of course it's an ugly looking image but then in certain screening exams like mammography or any other screening exam this false contour can be mistaken for a pathology which is not good so I'm just talking about in terms of the digitization let's take a single linear a to d converter to digitize luminant data from 0.1 optical density to an optical density of 4 actually you'd see such a thing in a mammogram because in mammograms you want to see the skin lines as well as the micro calcifications that are fairly dense it's a very high dynamic range and also the exposure is very high in mammogram mammography but it's at a lower kev so how many bits okay so essentially if I were using a linear a to d converter you know these calculations show that I would need the smallest step in a converter to be about a million so this would have this could not be a 16 bit converter this has to be a converter of 20 bits and that's quite impossible at the speeds that we are talking about so but you could use a log amp before the converter if you use a log amp today we don't use log amps because there are other issues you could use a 10 bit converter or you could use overlapping converters and do some curve fitting see and these calculations also valid when you're digitizing film where you're taking luminance and converting it to density in an imaging system and why is that important because some of the early mammography computer aided detection algorithms as we'll talk about used to digitize screen forms there's still systems like that and do try to find micro calcifications or speculations or you know extra ductile carcinomas which are important so LCD monitors as I said earlier are being used to display medical images and newer monitors have a resolution of 9 megapixels and we talked about off-axis viewing and management of imaging images on these monitors is quite a problem because what people like to do is to like to look at a whole bunch of images in the screen film situation you have a thing called alternators where people put a whole bunch of films and they you move racks of film up and down that's a very high bandwidth so what today people do is with fast storage they stage these images before a radiologist goes and looks at it so you can look at it fast but this was an issue until a few years ago before you had large bandwidth systems so this is a typical system it's a Kodak system which is a 7500 digital radiography system and it's a 750 thousand dollar system that's not why it's called 7500 but this system typically has a big detector here it's got an x-ray source it's got a table and under the table bucky where you could put other detectors and this detector moves up and down this is a Cadillac of a system and it uses a cesium iodide scintillator it's a panel made by a company called Trixel that's quite common so what are the advances in digital radiography retrofitable DR see today what happens in an existing digital radiography people had to replace a complete room you have to have the right type of generator to have the right detector by the time you replace the room it's a 750 thousand dollar system not today digital radiography systems are coming where you can retrofit it into existing rooms because there are about 50,000 conventional rooms still around in the united states and to and so people are interested in retrofitting at least half these rooms with retrofitable detectors the other is the portable DR whereas you take a detector and put it under a patient and then you can do a portable exam like in an ICU but these things have to be quite rugged because bariatric patients or patients who are very heavy can be about 350 pounds and the requirements for these detectors are that a 350 pound patient should be able to stand on the center not break it and notice it's a glass detector so you got to put a lot of stuff around it and then the new paradigms you'll talk about this are tomosynthesis and cone beam CT so tomosynthesis this is a very recent advance lot of it came out of mass general hospital and it talks about image visualizing it is about image visualization in mammography it uses CT reconstruction techniques to reduce anatomical noise let's talk about anatomical noise before that when you take an organ such as the breast okay it's compressed but you can only compress it to a certain point but what happens is when the x-rays are moving down there are layers of anatomy that interact with others and that's the image that you get and you got a small microcalcification or a speculated lesion and you can't see that many times because anatomy that has nothing to do with this blurs this so that's called anatomical noise and that was one of the early reasons why there wasn't a difference between screen firm systems and digital radiography systems even though these were more efficient so what was done by this fellow Nick Lawson who is with this company Hologic he was at mass general this is the breast here so he took x-ray images by moving the source around the detector so in CT you do reconstruction techniques that are based on equations that have to do with sums of linear attenuation coefficients but we do it a little differently so you shift the images to select planes and then you create tomograms and this is what happens these are early images these are a lot better today this was the original image and you look at the tomoplane one here you see you hardly see anything here tomoplane two three and see here there is here you see some micro calcifications that are very very visible but what happened see in the mammography people use typically two views there's a craniocaudal view that's from head to toe and the medial lateral view and people were thinking until recently that because you had tomosynthesis that one view was sufficient but it turns out that recently there were papers at RSNA that showed that if you had a lesion that was perpendicular you couldn't see it and also you needed to see axillary the axillary fossa because you a lot of infiltration of cancers occurs over there so it turns out that you do need both views in tomosynthesis but the advantage of this is use the same amount of dose you use less dose for each projection in order to get the total integrated dose which is about the same as a mammogram with a cc and an ml o view the only thing here is it takes more time in terms of compression so one has to worry about the comfort of the patient so this is a recent picture which actually showed a fast positive in mammography this was a conventional digital image and basically this is an ml o view because you see the this structure and also in mammography it's very important to come very close to the skin line that is not there and other detectors makes it complicated in terms of detector so if you look at it this looked very suspicious but if we looked at the tomosynthesis image this is on this is normal anatomy and it's not malignant and this was borrowed from a logic recently so let's talk about another advance it's called cone beam CT and it's got it's also an advance in mammography so here is a this is from Ning in University of Rochester and he took he just modified a CT scanner and instead of standard CT detectors he took a flat panel this a digital radiography detector that can go from go at high speed it's a 40 30 cb detector from variant medical systems that can take 30 frames a second was designed for fluoroscopy and if you look at this you take a mouse and you look at the memory tumor you can actually visualize it in a very pristine way because you have these elements these detector elements coming very close to each other this was some early work by dr. Ning and then today you put cone beam CTs on a C arm and if you are in a operating theater or in a surgery suite you can actually for image guided surgery you rotate the C arm and you can get a 3d reconstruction and you can do placement of images of let's say needles or biopsies or other things very accurately it's very important in brain surgery in neurosurgery that an image guided surgery so that you don't damage anatomy that you're not supposed to damage and this is another this is by dr. C. Woodson of Prince's Margaret hospital in Toronto where he did he came up with a system that looked at a cadaver head that was a previous thing that you saw now a later development this is the latest development cone beam CT for mammography what's happened is the patient is lying here and the and the breast is here as a pendant and then here's a flat panel detector here's a x-ray source and it rotates around the patient and so you can do a CT of the breast and this doesn't involve compression but it's immobilization but the interesting thing here is the source here is a tungsten source it's not a molybdenum source which is used in so this actually dr. Ning started a company called coning don't know how you got this name but and this is a system that's going to come out in a year and a half it's a cone beam system system for mammography you do one breast at a time at a time where the breast is in this cavity and this is the cone beam system on the inside it's a 10 second scan for breast and whereas you can get you can actually see lesions because that are otherwise obscured by anatomical noise because this is a true CT scan yes in visa v a standard CT a flat panel CT you get some scatter because the scatter reduction you got to have a grid and you can put a grid here that is an angular grid because it's a cone beam but typically scatter is not such a big issue we try to remove that in pre-processing and post-processing and CAD for mammography CAD is computer assisted detection originally what was done was we had screen film systems where the film was digitized and then detection was performed using CAD algorithms and today there are quite a few algorithms and FFDM by the way is full field digital mammography system okay that is the what I mean by FFDM I did not expand that and today you find actually CAD is efficacious for detection of microcalcifications but it's not efficacious for other things but then its sensitivity for microcalcifications is quite high 98 percent and for masses it's about 84 percent and for masses with microcalcifications it's 92 percent it's a good pre-screening thing or even a post you know diagnosis thing to see whether you miss something and it is still a very hot field and a lot of work is being done on it and there's a lot of controversy surrounding mammography with recent studies on it basically depending on whether somebody had dense press or non-dense press and this is an ongoing controversy in terms of what's efficacious so let's see what else can be done in with systems okay this can for example there are certain things in medical imaging where PET phogestron emission tomography would gives you physiological information and then anatomical information comes from MR or CT and really this is a single photon emission tomography it's nothing but a rotating gamma camera and so you register these and then you get a nice image which is an MR plus spect image where you're getting an anatomy as well as physiology if you go to Tata Cancer Hospital they have a PET CT and which is a PET scanner and CT on the same gantry and where as you see an extra cranial study of the thorax where the top row is a PET image the bottom row shows MR with a contour and the middle row shows image registration using MR and PET people have been doing this with CT but people are doing it with MR as well and if you want to detect PET CT detecting lung cancer the left top is a PET image of the thorax the right top shows the x-ray CT scan of the same the bottom image are because of registered images and actually the x-ray images that come out of the CT are used to correct for their attenuation in the PET scan as well we won't go into that at the moment I had and I gave a talk last year over here in July I did spend some time talking about it okay 3d medical imaging in telemedicine today you have high-speed broadband okay and what has happened is there's a company called terror recon which has come up with a board that finds itself in philips and semen systems where you could take 2d CT MR scans processed in a centralized server and you get a 3d model you get visualizations you get surfaces and then you take the 3d image you can actually segment the image before you send it or you can do it later and you send it by broadband to doctors at remote sites and then you see it's an asymmetric processing so it takes less time to actually look at the image than to create it and then you can look at it with laptops that's been happening and this is these are some images from the latest RSNA what are the benefits we talked about better patient visualization provides better and faster diagnosis and it saves on film cost today you don't need film except sometimes for people to archive and you may not need that either if you can archive things in a redundant fashion and basically you don't have mail latency and travel time and I can see in a future and I will talk about I'll spend the rest of the time after the talk in terms of the opportunities that are available in India in terms of 3d now I can presumably see once we start having faster connections or even reasonable connections that some of these images are sent overnight and staged and the diagnosis could be done offshore the time is coming so this is the end of my formal talk