Digital Imaging, an introduction

The Greeks had this concept of the atom. Keep splitting anything and eventually you will get down to the smallest unit that is particular to the thing you started with.

We regard the world as continuous but sometime use numbers to describe (or model) it. As soon as a number is used, an arbitrary decision has been made about a boundary. The more precise is a number, then the the closer we get to the way we normally perceive the world. This argument is the key to Zeno's paradox and theories of Chaos.

In digital radiography, we start with a number. It is the total number of X-ray photons that the patient may receive, in safety. We decide that number using technical and risk-benefit reasoning.

The total number of photons can be divided into spatial information and grey scale. Any digital image is a good analogy, since all digital images can be expressed in binary numbers in the same way that newspaper pictures are made of clusters of dots. Not all the number is available for pictorial information. There are some variations that we cannot predict which will reduce picture quality.

The beginner might be advised to examine assumptions about image quality,

1.	There is no guarantee that at any given time the output of X-ray photons from the Tube is constant. The use of heavy equipment in the same building or a nearby particle Physicist might affect the voltage.
1a.	Voltage variations may affect the reading cycle as the image is digitized. Once the image is a series of numbers this problem ceases.
2.	The photons are not always uniformly distributed in space. The tube geometry is optimised to a given voltage between cathode and anode.
3.	The silver particles in the film and the film chemistry are not uniform.
3a.	The compounds in the detector and their quantum efficiency may not be uniform in space or time.
3b.	There may unpredictable effects from the combination of the waveforms of the sharpening algorithm with the waveforms within the image. This is a very important feature in the analysis of modern digital systems. Diagnostic value of the final image has been improved by using a processing algorithm that is matched to the anatomical area under consideration. Since patients vary so much in their weight and habitus, such algorithm matching is not perfect.
4.	Patient factors of movement and X-ray scattering degrade accuracy.
4a.	Since digital image processing involves the amplification of low contrast features and increased Compton scattering from the higher Kilovoltage, there will be artifacts from x-rays that hit adjacent objects or adjacent tissue not intended to be in the final image. (remember that the processing assumes things about the source that may not be accurate).

There are similar assumptions about displays.

Return to start?

As soon as simplistic discussions on digital imaging commence, one soon gets embroiled in technicalities of contrast and resolution. The signal can be divided into various qualities. However you split it, the total number of information-containing photons is unchanged. The number of photons is then divided among the detector pixels (or the particles of silver that the photons may develop). Each pixel must have enough photons to separate out a range of grey scale.

The digitizing method will involve detectors of small size, but not at the expense of their efficiency. The quantum efficiency of some devices can be improved by increasing their size. Try to read a book in moonlight and you will see how the retina improves detection by recruiting detectors and lowering resolution. The number of detectors in a given space will be inevitably larger than the practical resolution that will result.

Incidentally, the appearance of the grid is deliberately chosen to remind you of the size-brightness relationships that may produce illusions.

Any detector of either location or grey scale must make a decision on the position of a boundary and produce a discrete number. In this CT scan, a line of Hounsfield numbers from one of the early head scans has been plotted. (Originally on a Sinclair Spectrum) It is very difficult to decide where the ventricles are. The retina solves this problem with surround inhibition.

If we extract as much picture information from our original number as is possible, then more detailed analysis will reveal only the random variations due to the above. Radiologists call this quantum mottle. At present, digital images contain less information than is in the distribution of photons that have passed through the patient. Compromises have been made to program models to allow the reproduction of diagnostically significant information.

The trade-off between contrast and resolution is well known in Art so that daubs of paint seen close-up become a boulevard, when viewed from a distance. Similar techniques can be used with digital images, provided the imaging system knows a little about the original object, the patient.

A seemingly random collection of squares becomes a recognisable feature and even simulates a curved surface in the same image at a lower resolution and in its context.

There have been numerous expensive failures in the implementation of the all-digital hospital. One of the more celebrated was hailed at the 1985 British Institute of Radiology congress.
One of the mistakes that were made in that project was consideration of what diagnostic content in a plain film really meant. Put simply, the problem was that, no matter how seductive the resemblence between an image and reality, diagnostic images are used by people to make diagnoses and discriminating between pathologies that resemble each other is what really matters.

Intensive care units have a more predictable and much smaller subset of diagnostic possibilities. Digital imaging has proved particularly helpful in this environment where improvement in imperfectly generated images, immediacy and closer supervision are fundamental to the existing clinical care.

The difference between diagnosis and teaching, or the demonstration of a new digital imaging system, is that in a teaching database one can artificially enhance contrast to give an impression of better resolution.

In a teaching database the same technical limitations affect the quality of the images but do not stop anyone from artificially enhancing their readability. In literature, it's called "Poetic Licence". The diagnostic answer is not unexpected here. The clinical context and radiological hints are already provided. The digital image for diagnosis has a much harder job, namely separating pathologies that may look alike in a patient without a known diagnosis.

The highest resolution of an object is best when its fine boundaries exactly correspond with the position of the boundaries of each detector unit. Artificially generated teaching images can be massaged to do just that. Technicians can investigate an imaging systems capacity to handle waveforms in the same way as one would investigate the quality of high fidelity sound equipment. In this context, the teaching database would be the equivalent of the electronic sound Synthesiser.

In this example of the unexpected, Each object in the right hand image is easier to recognise in its context.

The fact that 75% of all motor cycling accidents in the UK. are caused by motorists pulling out too soon is an indication that our visual system is very good at filling in the gaps in a received image. After all, the human visual system is set-up to recognise known structures from limited information. Our forebears would never have survived the African Savanna had they not not evolved a system of quick response to minimal stimulus. Ben Felsen called it the 'Aunt Minnie' effect (if it resembles Aunt Minnie then it must be Aunt Minnie).

Final, Final Warning

Beware of the fatal combination of the retirement of one or more powerful people and any large project in the Health Business.