Ewa Pietka 1 Introduction 783
Silesian University of 2 Background Removal 784
Technology 2 i Collimator-Caused Background 2.2 Removal of Background Outside Patient Boundary
3 Improvement of Visual Perception 788
3.1 Window/Level Correction 3.2 Histogram Modification
4 Image Orientation 789
4.1 Detection of the Rotation Angle 4.2 Image Rotation
5 On the Accuracy of Quantitative Measurements in Image Intensifier Systems 793
5.1 Geometrical Description of Pincushion Distortion 5.2 Surface Area Correction
5.3 Clinical Implementation
6 Implementation of Image Standardization Functions in HI-PACS 797
6.1 Industrial Standard for Image Format and Communication 6.2 Image Content Standardization Functions
7 Summary 799
Hospital Integrated Picture Archiving and Communication Systems (HI-PACS) have become multivendor and multiuser medical systems. Generally, there are two levels of standardization to be considered (Fig. 1). One level is required by the multivendor equipment implemented in PACS (i.e., acquisition stations, workstation, archive library). As a result of integration, PACS receive also data from other information systems working in the hospital (i.e., the Radiological Information System (RIS), the Hospital Information System (HIS), or the Digital Voice System (DVS)). Information systems deliver data to be viewed and analyzed at multivendor workstations. The second level is required by human subjects who address the system according to their professional duties. Among them there are radiologists, clinicians, technologists, and admission service managers. The format and content of the data varies and may include diagnostic images, medical reports, comments, or administrative reports. Thus, both levels of standardization have to be considered to make the entire system more effective in the clinical environment.
The first level of standardization is required to integrate all pieces of equipment through industrial data and image standards capable of handling information access and intersystem communication needs. The second level of standardization facilitates the medical requirements and use of the data. In this approach we concentrate on the image content adjustment to make images more readable and of better quality in preparation for medical diagnosis. This step also makes advanced image processing phase easier and permits some preprocessing steps not to be considered at the level of development of the methodology, which results in a computer added diagnosis. Although this group of standardization procedures is related to image information itself without referring to image format or intersystem communication, standardization functions are integrated with clinical PACS and installed at various sites of the system.
The first two sections of the chapter discuss improvement of X-ray image quality by the removal of image background caused by blocking of the collimator as well as background outside the patient boundary. Then, selected histogram modification techniques are shown for adjustment of anatomical structures under consideration. Section 4 discusses the image orientation problem and its correction. Since acquisition stations linked to PACS also deliver images mapped on image intensifier systems, Section 5 discusses the most important sources of distortions in quantitative measurements. Section 6 briefly discusses the industrial standard (yet the reader is asked to refer to other sources for details) and then gives some outlines of how to implement image standardization functions in clinical PACS.
Copyright © 2000 by Academic Press.
All rights of reproduction in any form reserved.
2 Background Removal
Various definitions of background have been introduced already. Usually it is described as an area of no importance attached to a region that is to be enhanced. Furthermore, very often it affects the visual image quality. In image standardization three various areas may be referred to as background. First consideration is given to the area outside the radiation field, caused by blocking of the collimator and resulting in white borders surrounding the radiation field. The other two types of background areas are located within the radiation field. Depending on following image analysis steps and expected result, the area referred to as background may change from one phase of analysis to another. The first phase of the image analysis (often a part of the preprocessing stage) usually concentrates on detection of the patient contour and its orientation correction. At this point the area outside the patient boundary is referred to as background and set to zero. Later phases of the image analysis, including segmentation and region of interest (ROI) extraction, may consider areas inside the patient boundary as background. For example, bone analysis in orthopedics or phalangeal and epiphyseal analysis in bone age assessment leads to considering soft tissue as background with respect to bony structures. Chest analysis also may require arms, mediastinum, and subdiaphragm to be referred to as background. The goal of many image processing functions is to suppress the background and/or enhance the foreground (i.e., diagnostically important regions) in order to increase the signal-to-noise ratio and/or extract features ofhigh discrimination power.
This section deals only with background to be removed by image standardization ( preprocessing) functions. Thus, removal of collimator-caused background and areas outside the patient contour are discussed. In the following sections two different regions are defined and later refer to as background. In Section 2.1. a region outside the radiation field caused by blocking of the collimator is called background, whereas in Section 2.2 an area outside the patient's boundary (within the radiation field) is referred to as background.
2.1 Collimator-Caused Background Glare in Image Diagnosis
Optimization of the amount of light exposing the viewer during the image reading procedure significantly affects the diagnostic performance, which is a result of information processing based on properties of the human visual system. Several sources of extraneous light may be pointed out, beginning with an increase in light reflection from the surroundings, which in turn increases the reflection of light from the radiograph itself. This phenomenon influences the perception of details in the radiograph. Then, an increase of the background level changes the adaptation level of the photo-receptors in the retina, shifting the visual sensitivity threshold. This process decreases the photoreceptors' response to a given stimulus increment and causes reduced contrast discrimin-ability. Involuntary eye movements have also been implicated as a source of decreased visual sensitivity. Finally, transparent areas within the field of view resulting in excessive light hitting the viewer's eyes increase glare, which causes two principal effects. First, the glare results in significant eyestrain and discomfort. Second, it decreases visual contrast sensitivity, which is proportional to the Weber ratio WR, where BB is the background brightness and RB is the region of interest brightness.
Glare also causes problems in digital projection radiography. Two conditions create glare in the soft-copy display; first, the surrounding light reflected from the display screen, and second, the light emitted by the screen, which results from parts of the image as well as from its periphery.
Extraneous surrounding light is limited by providing low ambient light in reading rooms. This reduces the reflection light to a level that does not shift the visual adaptation levels or alter perception to any appreciable degree. The second source, transparent background in the periphery, can be reduced by implementing software functions (as described later) to black out the borders. In return, this increases the visual quality of the image.
Whenever a collimator appears in the radiation field, transparent borders are generated within the field of view (Fig. 2). Their removal reduces the amount of unwanted light in the images during the soft-copy display as well as almost transparent borders on the film. Moreover, the removed
FIGURE 2 CR image with an unexposed background (white frame surrounding the image). (Courtesy of J. Zhang.)
background, without delivering any pertinent information, adversely affects observer performance.
Two major advantages are gained by implementing background removal in clinical PACS. First, it immediately provides lossless data compression — an important cost-effective parameter in image archive and communication. On display stations a more representative lookup table (LUT) pertinent only to the range of gray scales in the diagnostically important part of the image can been assigned. It also shortens the response time and permits a full-range image display on smaller monitors. Second, background removal already performed reduces effort while designing further image standardization steps as well as computer-aided image diagnosis for a certain type of image. It will improve the subsequent automatic image analysis, which will concentrate on the diagnostically important information with no background rather than the combination of both types of information.
Background removal becomes a challenging procedure to be incorporated in clinical PACS and performed automatically on all acquired images before sending them to the archive station. In case of failure, the software procedure may irreversibly destroy diagnostically important parts of the image (e.g., turn part of the lung area black). Only a repetition of the radiological procedure is able to undo this failure. All developed procedures decrease the accuracy of full background removal in order to ensure that no diagnostically valid part of the image is destroyed. The accuracy of developed methods ranges from 42 to 91% of full background removal and 86 to 99% for full and partial background removal.
One of the approaches with high accuracy of background removal has been developed by Zhang. Images may include pure background that does not overlap with any external objects or dense anatomical parts (Fig. 2), prostheses (Fig. 3a), or markers projected onto the background due to insufficient thickness of the collimator, high sensitivity of the image plate resulting in anatomical structures visible on the background (Fig. 4a), and odd-shaped abdominal collimators.
The first step of the algorithm analyzes the intensity distribution of the CR image background and determines the probability that a pixel belongs to the image background. The estimation is based on a background characteristic found in 1000 sectors marked on 50 randomly selected clinical radiographs. Within each sector a relationship between the average intensity of consecutive pixels yields a parameter called a background score. Assignment of each sector to background or diagnostic field permits the relation between background score and background probability to be defined.
Recognition of image background is defined by Zhang as the location of background edges. A gradient method is used in which the differentiation of image I yields a vector V x, y defined as
A set of all V x, y forms a gradient image. A pixel is selected when its p x, y exceeds an empirically determined threshold and the pixel itself is located close to the image border. The selection of pixels to be considered in further analysis is based on the background probability described earlier. Then, each pixel is subjected to the following background edge condition. If the difference between the average intensity of a pixel and its neighbor toward the image center is greater than that of the pixel with its neighbor toward the image edge, then this pixel is assumed to be on the edge.
In order to eliminate the failure of background removal, in which the diagnostically important part of the image would be erased, two additional conditions are imposed. First, pixels of low score are excluded. The second condition is based on an assumption that collimator edges are straight lines. Thus, the angle distribution curve of background pixels has four maxima. Location of these peaks corresponds to four edges.
Pixels with gradient values that fall within half width of the peak location are used to fit lines to be considered as background edges, The contribution of every pixel to fit the line is different with various weights defined as w
FIGURE 4 CR image with an unexposed background, (a) Original image with anatomical structures visible in the background; (b) background-removed image with enhancement of diagnostic field, (Courtesy of J, Zhang,)
where A and k are positive constants, and AO is the difference between the current pixel gradient angle and the maximum angle in the curve distribution angle,
Finally, yet a very important step in background removal is the estimation of reliability. The goodness-of-fit of already selected points is performed by applying the chi-square fitting technique. The second step is based on a comparison of two histograms. One is obtained from the original image, the second, from the image with no background. If in the diagnostically important part of the image the histogram remains unchanged, the background removal is accepted. Otherwise, the removal is ignored. Performance of the function is shown in Figs 3 and 4.
Removal of collimator-caused background is a modality-dependent as well as anatomy-dependent procedure. It is performed only for computed radiography (CR) images and digitized images. However, not all anatomies require its implementation. It is performed in pediatric radiology (for all anatomies) and in limb, hand, and head imaging (for adults). Collimator-caused background should not appear in adults chest and pelvis radiograms. Since only background (none of the anatomical structure) is subjected to the analysis, one procedure can handle all anatomies.
In this section the term background is referred to the image area outside the patient boundary, yet within the radiation field. In this area landmarks or labels with patient demographic data (name, birthdate, ID number, etc.) also may be found. It seems to be the most intuitively accepted and most often used definition of background in image processing systems. Selected approaches to the problem of background analysis, separation from the anatomical boundary, and removal are discussed later.
In most cases the procedure is anatomy-dependent and has to be developed separately for each anatomical structure. Mostly it is applied to anatomical structures clearly separated from the background. As examples, CR hand and limb images or mammograms and sectional images may be subjected to such functions. The background removal not only serves image viewing or image preprocessing functions, but is often used as a lossless compression method in order to decrease the size of an archived or transmitted image.
Various methods have already been presented, yet none of them yield results satisfying all anatomical structures. Two approaches are suggested in this section. One is based on a histogram analysis and another uses a dynamic thresholding technique, where the threshold is found by a statistical analysis of the background.
An image histogram is a gray-scale value distribution showing the frequency of occurrence of each gray-level value. For an image size of 1024 x 1024 x 8 bits, the abscissa ranges from 0 to 255; the total number of pixels is equal to 1024 x 1024.
Modification of original histograms very often is used in image enhancement procedures.
The histogram analysis is based on an assumption that the gray-scale values of foreground (anatomical structures) and background (outside the patient boundary) are distinguishable (Fig. 5a). This results in two peaks appearing on a histogram (Fig. 5b). Those peaks usually overlap, yet a minimum in between can be detected in order to separate both objects. After smoothing of the histogram, the threshold value can be determined either by locating the local minimum and maximum, or with statistical methods. This separates the foreground (white region in Fig. 5c) from the background (black region in Fig. 5c).
This approach fails in cases of nonuniformity of the background. This very rough assessment of the threshold very often cuts some parts of the anatomical structures, particularly the border areas between background and foreground. Parts of soft tissue have been cut in Fig. 5c.
In the dynamic thresholding approach the threshold is adjusted dynamically and depends on the current background value. In the first stage, a window of fixed size is scanned in the vertical direction and statistical parameters such as mean, variance, maximum gradient, and maximum value are computed. The four windows of lowest variance located in the image corners become candidates for a background area.
Then, the window with the highest mean value placed in the central part of the image is referred to as an area within the anatomical structure. Because of various landmarks and labels placed in the image periphery, location of the highest mean value window is limited to the central part of the image. The ratio of the highest and lowest mean values indicates the image quality and will be referred to as the quality control parameter. It may restrict the performance of this function.
A local threshold is defined using the mean and variance values of background windows. Then, a linear interpolation in vertical direction yields the threshold value for each row. The interpolation is also performed in the horizontal direction, yielding the threshold value for each column.
The thresholding procedure, performed separately in both directions, blacks out pixels below the threshold without changing those larger than the local threshold. If necessary, landmarks and labels are detected by searching the densest area of the histogram for a gray-scale value with the least frequency of occurrence. Morphological filtering is used to remove all small noisy elements in the background. An erosion function with a 3x3 pixel structuring element turns to zero all background elements equal to or smaller than the structuring element.
Both approaches to patient's background removal implemented in clinical images destroy parts of the diagnostic field. The area close to the patient contour is partially removed. This
FIGURE 5 Removal of background outside anatomical structures, (a) Original image; (b) histogram (arrow marks the threshold value); (c) thresholded image: anatomical structures remaining in the image are marked in white, prevents the methods from being used in an unsupervised implementation for clinical PACS. They can be applied at workstations in order to suppress the background and improve visual perception (as described in Section 3) or as a preprocessing function in computer-aided diagnosis. In both cases they are used interactively and do not irreversibly destroy the diagnostically important image regions.
Visual perception is an essential factor in medical diagnosis performed on soft-copy displays, Monitors do not provide a diagnostically accepted standard of display quality if no brightness and contrast adjustment is performed. This becomes of particular importance when digitized images are read. Image enhancement may be required for various modalities and anatomical structures.
The procedure may be preceded by a background removal function (see Section 2). Two approaches are discussed. First, a manual window/level adjustment (Section 3.1) is performed at a workstation, and the user is responsible for the quality and enhancement of the displayed image. Secondly, image enhancement parameters are found automatically, usually by means of a histogram modification technique (Section 3.2), and stored in the image header. At a workstation a user can easily switch from one set of parameters to another. Each set gives the enhancement of a different anatomical region (bony structure, soft tissue, lungs, etc.).
3.1 Window/Level Correction
In the window/level concept two parameters are defined. Window is referred to as the range of gray-scale values distributed over the entire dynamic range of the display monitor. A decrease of the window value increases the contrast in the display image, yet gray-scale values outside the window range are turned to black or white. The center of the interval is called the level value.
The window/level adjustment can be performed manually at the workstation or automatically. A manual shift of the upper and/or lower edge of the gray-scale level changes the window value. A more user-friendly adjustment uses a mouse or trackball. A vertical movement typically changes the window level, whereas a horizontal shift controls the level value. The gray-scale value adjustment can be performed in real time using a lookup table. The mapping is accomplished by defining a curve in which output gray levels are plotted against input gray levels.
A computerized window/level adjustment procedure first finds the minimum and maximum of a global image histogram. In order to suppress the extraneous gray-level values, 5% of values are cut off from both sides of the histogram. This means that the minimum gray-scale value is the 5% gray scale level of the cumulative histogram, whereas the maximum value is found at the 95% level. Minimum and maximum values define the window range and their average value yields the level. Window and level are used to generate the default lookup table for the image display.
The computerized approach to the window/level correction can be applied for single images (e.g., CR or digitized images) as well as for computed tomography (CT) or magnetic resonance (MR). If a single image is analyzed, a histogram of one image is found. For CT/MR images the entire set of images is used to calculate the histogram.
The histogram equalization technique described in Chapter 1 can be used to improve the appearance of the image. Figure 6 shows the result of histogram equalization performed on a CR chest image.
Another histogram method has been introduced to enhance specific anatomical structures. The goal of this preprocessing function is to create several piecewise-linear lookup tables to adjust the brightness and contrast of different tissue density. The procedure has been developed for CR chest images but also could be adapted for other anatomical structures. The first step is to analyze the image histogram to find key breakpoints that divide the image into three regions: background (outside the patient boundary, yet within the radiation field), radiographi-cally soft tissue (skin, muscle, fat, overexposed lungs), and radiographically dense tissue (mediastinum, subdiaphragm, underpenetrated lung). By applying different gains the contrast can be increased or reduced.
Based on this approach several lookup tables are created. Some of them enhance (at different levels) the radiographically dense tissue, others the radiographically soft tissue. One lookup table is created with no enhancement.
In clinical PACS the analysis is performed at the acquisition gateway and parameters are stored in the image header. At the time of display the enhancement level is selected manually by the user to improve the brightness and contrast of a certain gray-level region.
Image orientation becomes a very important issue in radiological systems in which a soft-copy display is used in daily clinical procedures. As a standard orientation, the position viewed by radiologists is considered. Typically two acquisition modalities may yield rotated images. First, a film digitization procedure may result in a nonstandardized image orientation when a film is placed in the wrong position. Yet clinical investigation has shown that scanned images are mostly at the correct position, and the orientation problem does not appear as a critical issue.
Computed radiography (CR) is another modality that yields misoriented images. A survey in pediatric radiology has found that between 35 and 40% of procedures are not performed with a conventional orientation. It is caused by the patient condition as well as the clinical environment. The CR cassette can be placed at various orientations to accommodate the examination condition.
Because of the position of the image plate, eight various orientations are considered (Fig. 7). They are divided into two major groups. The first group of orientations includes correct (Fig. 7a), upside-down (Fig. 7c), and rotated 90 clockwise (Fig. 7b) and counterclockwise (Fig. 7d). Flipping around the y-axis of the anteroposterior projection (Fig. 7e) gives an additional three possible orientations: upside-down and flipped (Fig. 7g), rotated 90 clockwise and flipped (Fig. 7f), and rotated 90 counterclockwise and flipped (Fig. 7h).
A turn to soft-copy display diagnosis makes the orientation correction a major problem. Since the orientation correction algorithm refers to the anatomical structure within the image, different functions must be applied for each anatomy (chest, pelvis, hand/wrist, etc.).
For chest radiographs, the orientation correction function handles various types of CR chest images: adults pos-terioanterior (PA) or anterioposterior (AP) and lateral projections, and pediatric PA (or AP) and lateral projections. Although all these are chest images, there are different problems that must be addressed in each of these types. Adult chest images are much closer to a standard image than
pediatric images in the sense that the mediastinum is usually centered within the image and the image itself contains only the thorax. The problem becomes much more difficult for pediatric images because the range of variation is much wider. The image may or may not include all or part of the head, the location of arms is random, and the area of exposed subdiaphragm differs from one image to another. Very often even the abdominal region is included, changing the location of lungs within the image. The image is usually not centered and is positioned at a random angle to the vertical axis of the image.
Was this article helpful?