Learning Objectives
- Understand what “good clinical image quality” means
- Learn about the three key technical IQ metrics (spatial resolution, contrast and noise), how these can be measured and the impact on Image Quality (IQ)
- Understand advantages and limitations of digital images
- Identify the contents of DICOM clinical image files and the typical size/archiving requirements
- Learn about the contrast limitations of the human visual system and the value of post-processing tools.
- Calculate the size of image files from given detector specs, data and image characteristics
- Apply the concept of Nyquist frequency to digital imaging problems.
- Overview of Information technologies (IT) in healthcare, including Picture Archive and Communication Systems (PACS) and Electronic Health Records (EHR).
There are many different imaging modalities:
Image Quality (IQ) is a general and subjective concept best described within the context of the specific task. An image with a good IQ has suitable characteristics for the intended use which could be screening, diagnostic, intervention or follow up.
Note: IQ does not mean aesthetically beautiful images!
For example in breast imaging, a high image quality enables detection and characterisation of:
Ideally with a high (100%) sensitivity (the ability to correctly identify the structures, the true positive rate) and a high (100%) specificity (the ability to correctly identify the structures without disease, the true negative rate).
$ Sensitivity = \frac{True \ Positive}{True \ Positive \ + \ False \ Negative} \times 100 $
$ Specificity = \frac{True \ Negative}{True \ Negative \ + \ False \ Positive} \times 100 $
Image quality is affected by information content, perception/interpretation and decisions by the observer:
Routine quality control aims to monitor equipment performance over time and compare it with a baseline/reference to ensure it adheres to the intended standards through the lifetime of the equipment.
Modalities involving ionising radiation require image quality to be complaint with ALARA/ALARP.
Analogue systems using film were sensitive to an upper and a low threshold. Too low a dose would result in the film being underexposed and too high a dose would result in the film being over exposed.
However digital systems (CR and DR) have wider dynamic range and are tolerant to sub-optimal exposure conditions. Therefore it is very hard for the operator to distinguish whether the machine is malfunctioning and potentially the patient could receive too little or too much radiation dose (dangerous!).
A digital image is an array of numbers assigned to each pixel or voxel. In a digital image the picture is broken down into discrete blocks. In a 2D system each block is termed a pixel (picture element) and in a 3D system each block is termed a voxel (volume element). A digital image is numerically described by:
The array size determines the sampling frequency (pixels/mm). The higher the sampling frequency the better the representation of the object detail.
The bit depth determines the number of possible values that can be assigned to a pixel. Quoted as the number of bits allocated to the image, so the simplest image would be 1-bit = 21 = 2 possible values = black & white.
| 1 bit | 1 binary digit |
| 2 nibble | 4 bits |
| 1 byte | 8 bits |
| 1 word | 2 bytes (generally) |
| 1 kilobyte | 210 = 1024 bytes |
| 1 megabyte | 220 = 1024 kbytes |
| 1 gigabyte | 230 = 1024 Mbytes |
The Human Visual System (HVS) is a little under 8-bit i.e. can distinguish ~200 Just Noticeable Differences (JND) in grey scale level. Medical imaging detectors and displays are typically 12-bit (i.e. 4096 grey levels) as post-processing tools manipulate and optimise the image for HVS.
Note: The representation of an object improves as the array size and bit depth are increased.
Image Size = 2300 x 1900 x 2 byte per pixel
Image Size = 8 740 000 bytes
Answer = 8.3 Mb
Digital images are normally viewed:
The same image at different window width and level settings show different information. Post processing may generate artefacts in the image (e.g. high level of edge enhancement may suggest that an implant is loose).
Lossy compression can make a file a lot smaller, however it is required by law that medical images have a lossless compression (to avoid any degradation in quality that could cause a change in diagnosis).
| Advantages of Digital Images | Disadvantages of Digital Images |
|---|---|
| Wide Dynamic range | Lower spatial resolution (Still may be adequate for clinical task) |
| Post processing capabilities | Initial cost can be high |
| Portability & telemedicine | Users have to monitor dose/patient exposure closely |
| Security & backup | |
| Less physical storage space required | |
| Advanced applications (CAD, Image subtraction, tomosynthesis, etc) | |
| Clean and safe processing |
Medical images are usually in the Digital Imaging and Communications in Medicine (DICOM) format. A DICOM file has two components:
All electronic detectors produce an analogue signal which varies continuously and which depends on the amount of radiation (or other form of energy) received by the detector. In most modern electronic imaging systems, the analogue system from the detector is transformed into a digital signal, that is a signal that has a discrete, rather than continuous values. During this transformation obviously some information is lost.
| Study | Archive capacity required (uncompressed Mb) |
|---|---|
| Chest X-ray (PA + L, 2 x 2 kby) | 20 |
| CT series (120 images, 512 x 512) | 64 |
| SPECT myocardial perfusion study (TI 201) | 1 |
| US study (60 images, 512 x 512) | 16 |
| Cardiac catheterisation | 450 - 3000 |
| Mammogram (screening) 2x CC + 2x MLO | 32 - 220 |
Spatial resolution, contrast and noise are the three key indicators of Image Quality. From these descriptors, Signal-to-Noise Ratio (SNR) and Contrast-to-Noise Ratio (CNR) can be derived. When measured under controlled conditions these can be very useful values.
SNR shows how many times stronger the signal is compared to the noise (signal variations). If all the sources of non-random noise can be removed than then the dominant source of noise is random (Poisson) distribution.
$ SNR = \frac{signal}{noise} = \frac{signal}{\sqrt{signal}} = \sqrt{signal} $
CNR is a useful metric in medical imaging as it allows us to quantify subtle variations in signal between objects and their surrounding background.
$ CNR = \frac{ \vert signal_{obj} \ - \ signal_{bkgd} \vert }{noise_{bkgd}} $
Its best to have a high photoelectric absorption and low Compton scatter. Important requirements of an imaging system is that is has a high signal detection efficiency with a high SNR and a high CNR.
An ideal detector would produce an exact representation (sharp response) of the object irrespective of the spatial frequency. However in reality the response is more curved. Spatial resolution affects the visibility of detail in an image and the ability to detect small structures close to each other. Poor spatial resolution of the imaging system shows a blur in the image. Decreasing the pixel size improves the spatial resolution at the cost of more noise as there are less photons per area (unless you increase the dose to compensate for this).
Nyquist-Shannon’s Sampling Theorem states if you have a signal that is perfectly band limited to a bandwidth of f0 (cycles/mm) then you can collect all the information there is in that signal by sampling it at discrete times, as long as your sample rate is greater than 2f0 (samples/mm)
For example, if the maximum frequency in the object is 2cycles/mm then the sampling must be done at least 4 cycles/mm.
$ N_F = \frac{1}{2 \times Pixel \ Pitch} $
Where the pixel pitch is the distance between two adjacent pixels.
Under-sampling occurs at a sample rate below the Nyquist rate. This leads to misrepresentation of the signal, loss of information and generation of artefacts.
| Imaging Modality | Pixel Size | Nyquist frequency lp/mmm |
|---|---|---|
| Mammography | > 0.080 m | 6.3 |
| General Radiography | > 0.143mm | 3.5 |
| Fluoroscopy | > 0.200mm | 2.5 |
Contrast key to detect subtle signals and is determined by the relationship between the magnitude of the signal and the magnitude of the fluctuations in the signal (noise). It depends on the composition and thickness of an object as well as the properties of the detector such as noise.
An ideal imaging system would:
However no such detector exists and noise is fashioned which prevents the visibility of small/low contrast details. Some sources of noise include:
Noise can be reduced, but never eliminated completely. CNR provides valuable data to investigate drops in Image Quality.
HIT has changed the way healthcare is provided. It holds great promise towards improving healthcare quality, safety and costs. Some examples of IT in healthcare:
There were some key milestones in the development of PACS:
PACS continue to develop, with technological advances making implementation similar and cheaper. Much current development focus on workflow and systems integration. At the moment, PACS typically comprises of:
May have additional networks to the other IT systems (HIS, PAS, RIS)
This is a record of important clinical information about the patient, and provides key performance indicators for the hospital or specialist unit (e.g. to support research, help planning new services):
The EHR can be created, managed an consulted by authorised providers and staff across more than one health care organisation. It can bring together information from current and past doctors, emergency facilities, school and workplace clinics, pharmacies, laboratories and medical imaging facilities.
The UK shows the biggest take-up of electronic health records in Europe. $2.1 billion (4% annual growth) was spent by the UK by the end of 2015 compared to $9.3 billion (7.1% annual growth) spent by the US.
The top 10 functions where doctors globally perceive a positive impact of EMR and HIE:
However there are some challenges of implementing EHR. Potential of EHRs meets problems of implementation as they could distract from doctor-patient relationships, wasting valuable time and driving up costs (costly to maintain).
What does “good IQ” means in the context of medical images?
A good IQ has suitable characteristics for the intended use which could be screening, diagnostic, intervention or follow up.
What factors that influence IQ?…and perceived IQ?
Answer coming soon
What are the 3 key technical descriptors of IQ?
Answer coming soon
What are the main sources of noise in X-ray imaging? And their causes?
Answer coming soon
How can CNR be measured? What affects it?
Answer coming soon
How can the performance of a medical monitor be assessed?
Answer coming soon
What data is contained in a DICOM file?
Answer coming soon
How does image matrix size and bit depth affect image quality?
Answer coming soon
What differences are expected between an 12bit and a 8bit image?
Answer coming soon
How does SNR relates with the number of photons used to produce an X- ray image for an ideal x-ray imaging system?
Answer coming soon
What is spatial resolution and how can it be improved for a digital system?
Answer coming soon
Discuss advantages and limitations of digital imaging systems?
Answer coming soon
What are the 2 main functions of an Electronic Health Record (EHR)?
Answer coming soon
Give examples of impact of EHR on patient and the healthcare system.
Answer coming soon
What is PACS?
Answer coming soon
How can PACS affect workflow in the imaging department?
Answer coming soon
Discuss key requirements of a hospital PACS?
Answer coming soon
How could IT systems support the management of adverse incidents in a hospital setting?
Answer coming soon
Discuss the introduction of IT technologies in healthcare and how they can bring benefits to patients and the healthcare system?
Answer coming soon
In the plane of the detector what spatial frequency can be recorded by a 512 x 512 pixel digital fluoroscopy system with 150mm x 150mm receptors?
Detector size = 150mm x 150mm
Matrix size = 512 pxls x 512pxls
Pixel pitch (d) = 150mm/512 = 0.293 mm
Nyquist Frequency (Nf) = 1/2.d = 1 /2(0.293) = 1.71 l p/mm
Solution: 1.71 lp/mm
A grayscale chest radiograph is 35cm x 29cm in area and was digitised with a sampling frequency that preserves the inherent spatial resolution in the image which is approximately 5 lp/mm (line pairs per millimetre). Each sample was digitised with 16 bits.
(a) Determine the image array size (in pixels)
The minimum pixel size to preserve the frequency is calculated using the Nyquist theorem:
Nf=1/2p … p=1/(2x5 lp/mm) = 0.1 mm
Image array size (pixels) = 350/0.1 x 290/0.1 = 3500 pxl x 2900 pxl
(b) Calculate the memory (in Megabytes) required to store a chest radiograph composed of an antero-posterior (AP) and a lateral (L) view of the chest (i.e. 2 images)
Memory required for one image
= 3500 x 2900 x 2 = 20 300 000 bytes / 1024 bytes/kbytes
= 19 824 kbytes / 1024 kbytes/Mbytes
= 19.3 Mbytes
Memory required for AP + Lateral ~39 Mbytes