British Journal of Radiology (2005) 78, 519-527
© 2005 British Institute of Radiology
doi: 10.1259/bjr/26039330
Effectiveness of flexible noise control image processing for digital portal images using computed radiography
S Yamada1,2 and
K Murase, MD, PhD2
1 Department of Radiology, Kurashiki Central Hospital, 1-1-1 Miwa, Kurashiki, Okayama 710-8602 and 2 Department of Medical Physics and Engineering, Division of Medical Technology and Science, Course of Health Science, Graduate School of Medicine, Osaka University, 1-7 Yamadaoka, Suita, Osaka 565-0871, Japan
 |
Abstract
|
|---|
The purpose of this study was to evaluate the clinical efficacy of flexible noise control (FNC) image processing in off-line computed radiography (CR) portal images. An observer study was designed to compare FNC with multiobjective frequency processing (MFP) in images acquired using a CR portal imaging device (CRPID). The image processing of MFP and FNC used the same data and required no additional irradiation of patients, and all images were printed on 10-bit grey-level dry laser film. Four radiation technologists and one oncologist served as observers and evaluated 40 sets of images for three different treatment sites: brain, lung and pelvis. Six to 10 anatomical landmarks were selected from each treatment site. Each observer was asked to rate each landmark in terms of its clinical visibility and rate the ease of making the pertinent verification in the MFP- and FNC-processed images. In the ratings of the visibility of landmarks and for the verification of treatment ports, FNC-processed images were more visible than MFP-processed images except for several landmarks in the anteroposterior (AP) pelvis such as the pubic symphysis. The visibility of landmarks in FNC-processed images was comparable with that in MFP-processed images. The verification of treatment ports using the CRPID with FNC was generally achievable. In conclusion, this study suggests that FNC is effective for image processing of CR portal images.
|
Introduction
|
|---|
Portal imaging is a widely used procedure in radiotherapy for verifying that a tumour has been irradiated and that the surrounding healthy tissue has been spared. It is important to irradiate the prescribed targets accurately, and if it is possible to use a verification system that gives high quality images and can be used frequently, the planned target volume can be minimized, providing high quality treatment. Until now, radiographic film has commonly been used for the verification images, but electronic portal imaging devices (EPID) have recently become widely used. In Japan, as well as EPID systems, computed radiography (CR) systems, by which good image quality can be obtained, are widespread. In this study, we examined techniques to improve the quality of portal images obtained by using CR.
The CR system was marketed as a diagnostic radiographic device by Fujifilm Medical Co., Ltd. (Tokyo, Japan) in 1983 [1]. There were only two methods for image processing in the initial CR apparatus, gradation processing and frequency processing [2, 3]. Gradation processing was developed to control the density of image signals, whereas frequency processing was based on the unsharp mask technique. Figure 1a
shows an unsharp masked sample image. Low frequency components are enhanced when the mask size is large, while high frequency components are enhanced when the mask size is small. Image distortion occurs at the edges of the radiation field by a combination of longitudinal and horizontal unsharp mask processing. To improve the accuracy, it was necessary to restrict the degree of enhancement of frequency processing [4]. This restriction greatly hinders attempts to improve the quality of CR portal images with a low density contrast.
View larger version (83K):
[in this window]
[in a new window]
|
Figure 1. Computed radiography (CR) portal images of the head. (a) Examples of an un-sharp mask image. Black-framed artefacts were observed around the irradiation field (black arrows) and around the scale (black arrows on the broken line). The skin boundary around the head was unclear (white arrows). (b) Examples of an multiobjective frequency processing (MFP) image. Black-framed artefacts were reduced, and the skin boundary became clear.
|
|
An image-processing algorithm with improved frequency processing was developed by Fujifilm Medical Co., Ltd. in 1998. This has been called multiobjective frequency processing (MFP). When MFP is used the degree of enhancement of signals with high-density contrast can be reduced, and that of signals with low-density contrast can be increased. Figure 1b
shows the improvements in an MFP image, when compared with the un-sharp mask image, where the black-framed artefacts seen in Figure 1a (indicated by the black arrows) have been reduced, and low density contrast (white arrows in Figure 1a) has been improved.
MFP is now established as the most suitable processing method for portal imaging in Japan. However, there is no established processing method for removing noise from images, and this has been recognized as a problem [5].
To address this problem, six variable image processing parameters [2, 3], that can be changed according to different circumstances, have been introduced. For example, one parameter may be changed to suppress selectively the enhancement of images in low-density areas that have a high level of noise. These parameters are listed in Appendix 1, together with a brief indication of which aspects of the processing they control. Another effective method of addressing the noise problem is to absorb backward scattered radiation efficiently by attaching a lead plate to the back surface of the cassette, and this has led to the development of a commercially available cassette for portal imaging [6]. However, these methods reduced effective signals which could detract from image quality in some regions, such as near to bone edges [7].
Recently, a new image-processing algorithm for CR has been developed by Fujifilm Medical Co., Ltd., called flexible noise control (FNC) [8]. This separates the noise and effective signals of linear structures, such as bone edges, by optimizing their different shapes, and was developed mainly to reduce the effects of noise in low-dose diagnostic radiography. There are four parameters for FNC and these are listed, together with a brief description of their purpose, in Appendix 1.
This study was designed to investigate the clinical efficacy of FNC processing, in comparison with MFP, for digital portal images obtained using CR.
|
Methods and materials
|
|---|
Images
The CR portal-imaging device used was an FCR-5000R (Fujifilm Medical Co., Ltd.), and a ST-V imaging plate with a 1 mm thick frontal metal screen cassette was used in the study. This cassette, with a rubber sheet containing 43% tungsten by volume, was placed in front of the imaging plate with a sponge on the back [9]. The CR system has an effective area of 350 mm x 350 mm for a 3520 x 3520 matrix size (0.1 mm pixel–1). The acquired image undergoes conventional processing, such as gradation processing and frequency processing [2, 3]. MFP and FNC image processing were carried out using a general-purpose personal computer running the image-processing software. The image can also be printed out life size using a dry laser printer (Fuji DP-L) with 10-bit grey levels. 40 sets of images for four treatment sites (brain, lung, anteroposterior (AP) pelvis, and lateral pelvis were selected. Six to 10 landmarks were selected as references, as in the study by Yin et al [10], for each treatment site. All images were obtained using a Mitsubishi ML-15MDX linear accelerator (Mitsubishi Electric Co., Ltd., Tokyo, Japan) operating at nominal beam energy of 4 MV photons with the system in linac-graphy (LG) mode. The dose rate in LG mode is a factor of 10 lower than that in therapy mode, so 1 monitor unit (MU) in LG mode is equivalent to approximately 1 mGy.
45 MU were required to produce a conventional double-exposure portal image (15 MU for the blocked field and 30 MU for the open field) for all treatment sites except the lateral pelvis, which required 90 MU: 30 MU for the blocked field and 60 MU for the open field. Image processing by both MFP and FNC used the same data.
Image processing
Multiobjective frequency processing (MFP)
MFP has two functions: a frequency enhancement function and a dynamic range compression function [2]. Figure 2 shows the steps involved in MFP. Band-pass filter images with various frequency components are produced by determining differences between the original and unsharp mask images (step b). Processing of each band-pass filter image for enhancement corresponding to the level of contrast is performed, and an image is produced by a combination of the enhanced images (step c). An MFP-processed image is obtained by superimposing the combined image onto the original image (step d).
View larger version (62K):
[in this window]
[in a new window]
|
Figure 2. Outlines of multiobjective frequency processing (MFP). The original image and several unsharp mask images were prepared (step a). Several band-pass filter images were produced (step b) by subtraction processing of unsharp mask images with a similar frequency band. A combined image was produced (step c) by multiplying each band-pass filter image by each enhancement coefficient. An MFP image was obtained (step d) by adding the original image to the combined image.
|
|
In enhancement processing corresponding to the level of contrast, the degree of enhancement of each band-pass signal is determined in the following way. If the input signal intensity is very much lower than the standard level, the output signal is produced by linear enhancement processing. If the input signal intensity is higher, processing with a non-linear coefficient that suppresses enhancement is performed. The enhancement in the periphery of the irradiation field, in which the contrast of portal imaging is sufficient, can be suppressed by MFP.
Dynamic range compression (DRC) widens the visible range. It compresses the dynamic ranges in areas with both low and high densities while maintaining the density contrast of images in intermediate ranges. DRC was used in conventional methods, but processing was restricted to either white areas or black areas. In MFP, DRC in both areas is possible, and so this processing method is now the most widely used for CR portal imaging in Japan.
The MFP parameters for portal imaging used for the observer study are briefly described in Appendix 1 (Table A1). These parameters are recommended for portal imaging in Japan by the manufacturers.
Flexible noise control processing (FNC)
FNC suppresses image noise whilst retaining diagnostically important contrast in anatomical structures. Figure 3 shows the steps involved in FNC processing. Band-pass filter images with various frequency components are produced (step b) by determining differences between the original and unsharp mask images (step a), as in MFP. Subsequently, noise components are extracted from each band-pass filter image. In the extraction of noise components, blur signals that exist in the periphery of linear structures are suppressed from linear structures in the direction of the lines. For the signals from dotted structures, those with low contrast below a fixed threshold are regarded as noise, and are suppressed. Signals with high contrast over a fixed threshold are maintained. Differences between images with only signal components extracted from the dotted and linear structures and the band-pass filter image are determined, providing an image with noise components alone. This processing is performed in each band-pass filter image. An image is produced (step d) by superimposing the images with only noise components in each frequency band. A noise-suppressed image is obtained (step e) by subtracting the combined noise components image from the original image. Routinely, the combined noise component image is subjected to MFP as described above.
View larger version (74K):
[in this window]
[in a new window]
|
Figure 3. Outlines of flexible noise control processing (FNC). Several band-pass filter images were produced (step b) using several unsharp mask images, as in multiobjective frequency processing (MFP). Subsequently, images with extracted noise components alone were produced (step c), and an image with combined noise components was produced (step d). An FNC-processed image was obtained (step e) by subtracting the image with combined noise components from the original image. Routinely, the FNC-processed image was regarded as the original image, and was subjected to MFP.
|
|
In FNC processing it is necessary to define the noise level in order to distinguish between noise and signals from linear and dotted structures. In conventional diagnostic radiology the noise level is low for relatively high X-ray doses, but increases as the dose decreases, and hence, for FNC processing in these circumstances, the noise level is defined for each incident dose. Noise suppression is then a function of the dose. In portal imaging, since noise will occur at any X-ray dose, we have established a method whereby noise can be suppressed irrespective of the incident dose. Table A1 (Appendix 1) also shows the parameters used for FNC processing, and again these parameters are briefly described in Appendix 1. These parameters were not optimized to improve the quality of portal images, but are the minimum parameters necessary in order to examine whether further improvement of the portal image quality can be obtained by the most commonly used MFP with additional FNC processing.
Observer study setup
Four treatment sites (brain, lung, AP pelvis, and lateral pelvis) with 40 sets of images (10 sets per site) were selected for the study. Each set was placed on a viewing box. Six to 10 landmarks were chosen for each treatment site. Of the observers, A and B were radiological technologists with more than 5 years of radiotherapy experience. Observers C and D were technologists with less than 3 years of radiotherapy experience but more than 5 years of experience in diagnosis, and observer E was an experienced radiation oncologist. With unlimited viewing time, each observer evaluated each set of images on the viewing box, and provided ratings in response to two questions.
The first question asked for a direct comparison of the MFP and FNC processed portal images of each landmark:
Question 1: Comparing the MFP and FNC-processed images, is the quality of the FNC-processed images higher than that of the MFP images? FNC processing is (1) much better, (2) slightly better, (3) comparable, (4) slightly worse, or (5) much worse.
The second question asked for a direct comparison of the ease of the overall verification of MFP and FNC processed portal images:
Question 2: Comparing the MFP and FNC-processed images, is the verification of the treatment areas easier on the FNC-processed images compared with that on the MFP images (overall impression)? The FNC processing is (1) much easier, (2) easier, (3) the same, (4) harder, or (5) much harder.
To summarize the data, the ratings were treated as quantitative scores, and the means were computed for each observer for each question. For question 1 the effects of FNC processing is indicated as better (score<3), comparable (score=3), and worse (score>3).
Question 2 data were used to compare the proportions of rates that were "much easier", "easier", and "the same" (score<4) for the ease of verification. These ratings indicate the possibility that FNC-processed images can be clinically used instead of MFP images. These questions and the selected landmarks were determined following the method of Yin et al [10] who conducted an observer study on the direct comparison of clinical efficacy.
|
Results and discussion
|
|---|
Figure 4 shows examples of MFP-processed images on the left and corresponding FNC-processed images on the right. Figure 5 is similar but shows parts of the images after magnification. The results of the observer study in respect of the image quality for each landmark are shown in Table 1. In the brain (Figure 4a and 4b), the image quality in all anatomical parameters other than peripheral skin was judged to be improved by FNC processing. The contrast on the peripheral skin was adequate on the MFP images and this is why a high proportion was judged to be comparable. Figure 5a and 5b show the pituitary fossa region after magnification. The quality of the FNC-processed image of the pituitary fossa was judged to be better in 60% of cases.
View larger version (92K):
[in this window]
[in a new window]
|
Figure 4. Sample images of multiobjective frequency processing (MFP) and flexible noise control (FNC). The upper panels (a, b) show the head, the middle panels (c, d) show the chest, and the lower panels. (e, f) show the lateral pelvis. The left panels (a, c, e) are MFP images, and the right panels (b, d, f) are FNC images.
|
|
View larger version (96K):
[in this window]
[in a new window]
|
Figure 5. Parts of the multiobjective frequency processing (MFP) and flexible noise control (FNC) images after magnification. The upper panels (a, b) show the pituitary fossa, the middle panels (c, d) show the lung apex, and the lower panels (e, f) show the pubic symphysis. The left panels (a, c, e) are MFP images, and the right panels (b, d, f) are FNC images.
|
|
View this table:
[in this window]
[in a new window]
|
Table 1. Proportion of images for which flexible noise control (FNC) was rated "better", "comparable" or "worse" than multi-objective frequency processing (MFP), by site and landmark
|
|
In the lung (Figure 4c and 4d), high scores for "better" were obtained with the anatomical parameters of lung space and lung apex (Figure 5c and 5d). However, in the ribs, the rates of "better" and "worse" were 0.42 and 0.38, respectively, indicating that opinion was almost equally divided. The image quality of regions with low frequency signals, such as the bone edges of ribs, was improved by FNC processing, but the image quality of regions with high frequency signals, such as the trabeculae of ribs, was not improved because signals were attenuated together with the noise. The divided opinion may, therefore, have been because different observers focused on different features.
In the AP pelvis, the rating of "better" generally had a high proportion showing a preference for FNC processing. In this study, images of prostate patients were mainly used. For anatomical landmarks in high-density areas of the irradiation field, such as the pubic symphysis (Figure 5e and 5f) and iliopectineal line, the comparable ratings were 0.56 and 0.40, respectively, indicating no particular improvement by FNC processing. In the lateral pelvis, the usefulness of FNC processing was observed in the sacrum, in which a high level of noise was present in MFP images (Figure 4e and 4f).
Table 2 summarizes the overall evaluation of each treatment site. The rates of "much easier", "easier", and "the same" (score <4) in question 2 are shown for each observer. These ratings indicate that FNC-processed images may be used clinically instead of MFP images. Observers A, B, and E with significant experience of radiotherapy answered that the verification of treatment areas was easier on the FNC-processed images. However, observers C and D with longer experience of diagnostic radiography than radiotherapy indicated higher rates of difficulty in evaluating the treatment areas on FNC-processed images than those encountered by the other observers.
View this table:
[in this window]
[in a new window]
|
Table 2. Proportion of images for which flexible noise control (FNC) was rated "much easier", "easier", and "the same" for the ease of making a verification compared with that with multiobjective frequency processing (MFP), by site and reader
|
|
The large differences in evaluating the effectiveness of FNC processing between the observers were not considered to have been caused by differences in the experience of radiotherapy, but the observers with long experience in diagnostic radiology considered that CR portal imaging was diagnostically important and were uncomfortable when diagnostic information was diminished. However, for treatment verification purposes the ability to recognize outlines is generally considered to be more important, and so the processing in this study was primarily intended to enhance the outlines of bones and tumours. If the outlining information is adequate then images that retain more diagnostic information may be regarded as better. The observer study therefore, becomes very subjective and we must clarify definitions to improve image quality.
The parameters of MFP images used in this study were adjusted to optimize the image quality by MFP alone, taking the level of noise on the images into consideration. In future studies, variable MFP parameters should be adjusted to show each anatomical parameter with the highest signal intensity without taking noise distributed on the image into consideration, and then FNC processing should be performed, which will markedly improve the image quality of the AP pelvis, where the contrast is low.
In this study, FNC processing at a certain level of suppression was performed in all density areas on the images. However, the effect of FNC processing on the improvement of the image quality was poor in high-density areas, and therefore, it would be better to select variable parameters by which the degree of noise-suppression is low in these areas. With regard to the risk of suppression of useful signals by FNC processing, it would also be useful to evaluate the threshold for the distinction between signals and noise.
|
Conclusion
|
|---|
Where the use of portal imaging is for treatment verification purposes, additional FNC processing, which provides clear bone edge signals, is considered worthwhile as the image quality is increased in the lung airspace and bone edges. The performance of FNC in the brain was especially good.
We hope that the CR system will become more widely used and replace the film method, and a future aim is to develop a portal imaging system that provides image quality equivalent to the CR system and convenience similar to the EPID.
|
Appendix 1
|
|---|
Parameters used in the two processing methods. The six parameters used in multiobjective frequency processing are:
1. Multi-frequency balance type (MRB)
MRB determines the frequency characteristics of enhanced images. Six types are available. Type A mainly enhances lower frequency and type F, higher frequency. In this study, type C was used.
2. Multi-frequency enhancing type (MRT)
MRT is a parameter of the enhancement factor function. MRT determines the shape of the function when the maximum degree of enhancement is normalized to 1.0. Eleven types are available. In this study, type F was used. With type F, the maximum degree of enhancement is performed irrespective of the image density.
3. Degree of multi-frequency enhancement (MRE)
MRE determines the degree of enhancement. Value ranges from 0.0 to 9.9 and 10 to 16.
4. Multi-DRC balance type (MDB)
MDB determines the frequency characteristics of smoothed images. Seven balance types are available. With type G, it is possible to obtain smoothed images with the best edge images.
5. Multi-DRC enhancing type (MDT)
18 types are available. With type P both high-density areas and low-density areas were appropriately corrected.
6. Degree of multi-DRC enhancement (MDE)
MDE determines the degree of enhancement. Value ranges from 0.0 to 1.0.
|
Parameters used for FNC processing
|
|---|
The four parameters used in FNC processing are:
1. Filter control type of FNC (FFC)
FFC determines the degree of noise-suppressive processing depending on the radiation dose. FFC has modes A and M. With mode A, noise-suppressive processing becomes higher with decreases in the dose. This was intended for use in diagnostic radiography. With mode M, noise-suppressive processing is performed at a fixed level irrespective of the dose. This was developed for CR portal imaging. In this study, mode M was used.
2. Balance type of FNC (FNB)
FNB determines the degree of noise-suppressive processing in relation to frequency components. FNB has modes A to F. With mode A, noise is suppressed in all ranges from low-frequency components to high-frequency components. With modes in alphabetical order, noise-suppressive processing is performed in higher frequency components. In this study, mode C was used. With mode C, noise suppression is performed in a frequency range of over 0.4 cycles mm–1.
3. Type of FNC (FNT)
FNT determines the degree of noise-suppressive processing in relation to the image density. FNT has modes A, B, and C. With mode A, noise-suppressive processing is constant irrespective of the image density. With mode B, the degree of noise-suppressive processing becomes low with increasing density, while with mode C, it becomes high with decreasing density. In this study, mode A was used.
4. Enhancement of FNC (FNE)
FNE determines the degree of noise-suppressive processing in the range of 0.0–1.0. Noise-suppressive processing is highest at 1.0, and processing is not performed at 0.0. In this study, FNE was performed at 0.7.
|
Acknowledgments
|
|---|
We sincerely thank Mr Ryoji Sasada and other staff of Fujifilm Medical Co., Ltd. in Miyanodai, Japan.
Received for publication December 12, 2003.
Revision received November 18, 2004.
Accepted for publication January 6, 2005.
|
References
|
|---|
- Wilenzick RM, Merritt CR, Balter S. Megavoltage portal films using computed radiographic imaging with photo stimulable phosphors. Med Phys 1987;14:389–92.[CrossRef][Medline]
- Iwasaki N. Image processing. In: Iwasaki N. Fuji computed radiography. Japan: Fujifilm Medical Co., Ltd Press, 2000:29–51.
- Ishida M. Image processing. In: Tateno Y, Iinuma T, Takano M. Computed radiography. Japan: Springer-Verlag, 1987:25–30.
- Yabutani T, Endo Y, Suzuki S, Kato H, Kitazawa H, Sawada T. Effect of frequency processing on portal imaging in computed radiography. Jpn J Radiol Technol 1997;53:380–5.
- Yamada S, Murase K. Analysis of frequency components of CR portal images and application to multi-objective frequency processing. Jpn J Radiol Technol 2003;59:864–71.
- Sasagaki M, Matsumoto M, Mori Y. CR portal imaging –A linac graphy using storage phosphor imaging systems. Jpn J Radiol Technol 1992;48:984–90.
- Sato H, Yokozawa M, Nishizawa M. Study of intensifying metallic plate for linac graphy using computed radiography. Jpn J Radiol Technol 1997;53:1479–86.
- Iwasaki N. Flexible noise control. Fuji Medical Review, 2004;12:25–32.
- Sato H, Yokozawa M, Tanaka K, Sakai H. Study of cassette system for linacography using computed radiography: application of a heavy metallic sheet to the metallic plate. Jpn J Radiol Technol 1999;55:198–204.
- Yin FF, Rubin P, Schell MC, Wynn R, Raubertas RF, Uschold G, et al. An observer study for direct comparison of clinical efficacy of electronic to film portal images. Int J Radiat Oncol Biol Phys 1996;35:985–91.[Medline]
Top
Abstract
Introduction
