| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
From the Lions Eye Institute, Centre for Ophthalmology and Visual Science, The University of Western Australia, Nedlands, Australia.
 |
Abstract |
|---|
 Top Abstract IntroductionMethodsResultsDiscussionConclusionsReferences |
|---|
METHODS. JPEG (Joint Photographic Experts Group) and Wavelet image compression techniques were applied in five different levels to 11 eyes with subtle retinal abnormalities and to 4 normal eyes. Image quality was assessed by four different methods: calculation of the root mean square (RMS) error between the original and compressed image, determining the level of arteriole branching, identification of retinal abnormalities by experienced observers, and a subjective assessment of overall image quality. To verify the techniques used and findings, a second set of retinal images was assessed by calculation of RMS error and overall image quality.
RESULTS. Plots and tabulations of the data as a function of the final image size showed that when the original image size of 1.5 MB was reduced to 29 KB using JPEG compression, there was no serious degradation in quality. The smallest Wavelet compressed images in this study (15 KB) were generally still of acceptable quality.
CONCLUSIONS. For situations where digital image transmission time and costs should be minimized, Wavelet image compression to 15 KB is recommended, although there is a slight cost of computational time. Where computational time should be minimized, and to remain compatible with other imaging systems, the use of JPEG compression to 29 KB is an excellent alternative.
|
Introduction |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionConclusionsReferences |
|---|
One of telemedicine’s main attractions lies in the ability to provide specialist medical care to areas that are underserved, particularly those located remotely from major population centers. For example, the state of Western Australia (land area 2,500,000 km2) has a population of 1.9 million, 70% of whom are located in the lower southwest corner. Because all ophthalmic specialists are located in this area, special trips are made to cover the vast remaining areas. This isolation is worsened by high rates of diabetes, cataract, trauma, and endemic trachoma in the aboriginal population.3
Tele-ophthalmology has the potential to improve the accessibility of people in remote areas to specialist ophthalmic care, and in turn to help fight preventable blindness.4 It can also have a large impact on the costs and necessity of transporting patients to regional centers. As a primary screening tool, tele-ophthalmology also has a role in identifying patients needing nonurgent treatment. In this way, the expense of sending ophthalmic teams to remote, isolated, and sparsely populated areas can be reduced.
The essential parts of a store-and-forward telemedicine system include good quality data recording equipment and effective communications systems operated by trained health personnel, and a data archiving and viewing system accessible by specialist medical personnel. In the case of ophthalmology, where ocular imaging plays a significant role in clinical diagnosis, good quality digital images of the retina and external eye form the key part of the system.
Although the technical issues of image capture, digitizing, and transmission can easily be addressed, two factors that are linked can make the whole tele-ophthalmology system ineffective. Communication systems in remote areas are often of low quality and in some cases nonexistent. Satellite telephones can be used, but affordable systems suffer from low data transfer rates and are expensive to operate. Furthermore, digitized ocular images require substantial storage space and take a long time to transmit. A high-quality digital image can reach a size of 1.5 MB or greater. Over a 9600 bauds/sec modem line, this image can take at least 25 minutes to be transmitted,5 which becomes important when a large number of images are to be transmitted.
Image compression is common in the transmission of images. JPEG (Joint Photographic Experts Group) compression, the most common image format used on the Internet, is also used for medical imaging,6 7 8 although Wavelet image compression has also been investigated.9 10 To achieve an appreciable reduction in image size (i.e., more than 1:4) some loss of information and consequently some degradation of image quality must be expected.
JPEG image compression breaks the image into blocks of 8 by 8 pixels and converts these blocks subsequently into spatial frequency components. A sampling is made of this frequency domain information (in a step called quantization) by closely preserving the low-frequency components and approximating the high-frequency components. The amount of information that is discarded determines the amount of compression. A coding process compresses the remaining frequency coefficients. The decompressing process reverses these steps. The effects of compression can be seen at high compression levels when "blocking artifacts" become evident (see Fig. 1 ).
|
A number of studies have been conducted on medical image compression. Bittorf et al.11 assessed compressed images of skin lesions, concluding that images needed to be at least 768 x 512 pixels with 24-bit color resolution (i.e., >1 MB images) to be suitable for diagnosis. Persons et al.7 found that low-contrast objects in the images still remain visible after JPEG and Wavelet image compression of magnetic resonance imaging and computed tomography images, although fine and irregular details are easily degraded. Martin et al.6 compressed 2-, 3-, and 4-MB fundus images and found that a compression ratio of 1:24 still produced images of diagnostic quality.
A number of approaches can be taken to find the optimum level of image compression. Some studies determine the compression level at which changes from the original image are first noticed ("just-noticeable-difference"). In one study the compression ratio limit for chest radiograph images is shown to be 1:6.12 Erickson et al.10 rated the appearance of structures on chest radiographs, finding that a Wavelet image compression ratio of 1:40 produced images indistinguishable from the original images. Another approach is to determine at what compression level the abnormalities become indistinguishable or the highest compression level that is clinically acceptable. In a previous study we found that a JPEG compression ratio of approximately 1:30 (approximately 20–30 KB images) can be achieved without affecting the ability of the ophthalmologist to identify these abnormalities.13 Kim14 used JPEG compression for 900-KB gastrointestinal color images and showed that compressed images approximately 20 KB in size were still acceptable.
In the present study we continued our investigation to determine the effects of digital image compression on various types of retinal images and the level of compression tolerable. Of interest was the comparison of JPEG and Wavelet image compression techniques.
|
Methods |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionConclusionsReferences |
|---|
To test the limits of image compression, eyes with subtle abnormalities were selected; in most cases these abnormalities were of low clinical significance, requiring only to be noted for future observation. The abnormalities included small nerve fiber layer and macular hemorrhages, macular and peripheral drusen, and cotton wool spots. Four of the 15 eyes were normal controls, taken from the same set of eyes. Images with extreme abnormalities were not used because some preliminary investigations indicated that these images could be compressed to well over 1:300 using JPEG, and abnormalities in the retina could still be detected.
The 35-mm slides were digitized with a Polaroid SprintScan35 (Cambridge, MA) at a resolution of 675 dots/inch with 24-bit color, resulting in a file size of 1.5 MB (752 x 680 pixels). This produced a high-quality digital image, without making the image size too large to be unmanageable. The images were stored as TIFF (Tagged Interchange File Format) files. The images were all compressed to five different levels using JPEG and Wavelet algorithms. The goal was to produce some compressed images in which the image quality was too low to be useful for assessment.
JPEG compression was performed by an algorithm developed from standards set by the Joint Photographic Experts Group (http://www.jpeg.org/). The JPEG algorithm is also available in many imaging software programs, although there appears to be a variation among some of these in the definition of the compression level. The compression level is determined by selecting a quality (Q) value, and the exact compressed file size cannot be predicted. Q values of 20, 40, 60, 80, and 100 were used, which resulted in images that were approximately 350 KB, 49 KB, 29 KB, 21 KB, and 14 KB, respectively, in size.
The Wavelet compression was applied by using a baseline wavelet transform coder with an Antonini filter15 in custom written software. The resultant file size after Wavelet compression could be predicted; compression ratios were selected to produce files equivalent to the JPEG compressed images, but still using a "round" compression ratio. The compression ratios were 1:5, 1:30, 1:50, 1:70, and 1:100, resulting in image sizes of approximately 300 KB, 52 KB, 31 KB, 22 KB, and 15 KB, respectively.
After compression, each image was decompressed and saved as a 752 by 680 pixel TIFF image to ensure that each image was loaded for viewing at the same rate.
Four methods were used to assess the quality of the 165 images:
Objective Assessment by RMS Error
An objective method of measuring image fidelity was obtained by
calculating the root-mean-square (RMS) error between the original and
compressed images.16
This calculates the sum of the
differences between each pixel value in the original image and the
corresponding pixel in the compressed image. Each digital image
contains three color channels (red, green and blue; or RGB); RMS error
was calculated for each channel.
Vessel Branching
A semiobjective method of assessment involved observation of blood
vessel branching.17
Images were assessed by displaying
them on a computer monitor. Two assessors working together identified a
retinal artery on the highest quality image, noting the number of
branches of this vessel that were visible. A branch was defined where
the distal branch of the artery had a decreased diameter compared with
the prebranch vessel. The successive compressions of this image were
then displayed, with the assessors noting the number of branches that
continued to be visible. Grading of image quality was arbitrarily
determined by the following: four branches visible, excellent; three,
good; two, acceptable; and one, poor or unacceptable. Retinal arteries
were selected because they are generally thinner and are less
contrasted than the veins.
Ophthalmic Assessment
All compressed images and original images were displayed as TIFF
images in a random order on a 17-inch computer monitor (dot pitch 0.28
mm) set to 1024 by 768 pixels. Each image was shown with no zoom and
filled most of the screen. The same monitor was used by each assessor.
Ophthalmologists experienced in mass screenings were asked to note
their observations from all 90 images (6 of each eye); they were not
told the number of normal subjects, or the abnormalities that they
could expect to see. Four ophthalmologists assessed the JPEG compressed
images and three the Wavelet compressed images.
Sensitivity, specificity, and kappa agreement were calculated to summarize the assessments of the ophthalmologists. Kappa indices18 were used for agreement between the observations from the compressed images and the gold standard assessments. Values of kappa of 0.6 or over are generally taken as indicating good agreement.
Subjective Assessment of Image Quality
While assessing the images, the assessors were also asked to grade
the image quality as being good (image degradation not evident),
acceptable (image degradation evident, but still able to make an
assessment), or poor (quality not sufficient from which to make an
assessment).
After these images were assessed, a second set of images of eight eyes was obtained to determine whether compression also affected other types of images in a similar manner. These images, showing mild to gross retinal abnormalities, were digitized as described above and compressed to the same compression ratios. They were then analyzed by calculating the RMS error and by qualitatively assessing image quality: good, acceptable, or poor.
|
Results |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionConclusionsReferences |
|---|
Compression ratios to approximately the same file size were obtained for each right and left pair: 55 KB, 24 KB, and 15 KB. The blocking effect can be seen on the lower two JPEG compressed images, particularly in the lowest images. However, most of the details of the hemorrhage and the blood vessels are retained, despite their taking on a sheared appearance. The Wavelet compressed images on the right appear fuzzy and out of focus; the vessels start to blend into the surroundings to the extent that for the highest compression the hemorrhage starts to become indistinguishable from the neighboring blood vessels. In both cases, two small reddish dots near the top right-hand corner of the image disappear with progressive compression. In Figure 2 , fuller views of the most compressed images are shown; in both cases the hemorrhage can be seen clearly. However, the blocking effect of the JPEG image makes the Wavelet compressed image a more attractive image subjectively, despite being rather fuzzy.
|
|
|
|
|
|
|
|
|
Discussion |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionConclusionsReferences |
|---|
The results of the various methods of assessment of the quality of compressed retinal images are relatively consistent. They show that 1.5-MB images can be compressed to at least 29 KB (compression ratio 1:52) for JPEG compression and 22 KB (1:68) for Wavelet compression before there is a loss in image quality.
Although there are only 15 images in the study, the low standard deviation of the objective assessments (4.6% and 8.3% of the mean RMS error for JPEG and Wavelet compressed images, respectively) suggest that this is sufficient.20 Future studies with more images will confirm the power of the methods.
Compression ratios on their own should be treated carefully because they do not reveal the original size of the image. For example, Martin et al.6 found they could compress fundus images to 1:24 but started with images of 2, 3, and 4 MB. This means the final image sizes would have been 83 KB, 125 KB, and 166 KB, respectively. Therefore, it is probably better to use uncompressed and compressed image sizes to compare image compression.
One other, and also brief, report has been found of ophthalmic image compression. Anagnoste and colleagues21 compressed 12 color fundus and fluorescein angiogram images using JPEG format and assessed subjectively the projected images. They found that compression ratios of 1:28 and 1:12, respectively, produce images in which compression was not noticeable. However, the original image size was not provided.
Comment should also be made on the semiobjective and subjective methods used. Observation of the blood vessel branching was relatively straightforward. However, initial attempts displayed the images in a random order. Although the results still showed a trend similar to that presented in Tables 1 and 3 , there was some variation because the same vessel was not always tracked by the assessors. Smaller original vessel diameters degenerate sooner than larger vessels with compression.
The use of ophthalmologists to record retinal abnormalities was the most time-consuming; 90 images could generally be assessed in about 1 hour. Although display order of the images was randomized, the assessors were still in many cases able to recognize images they had previously seen. This section of the study could have been changed to overcome this. The number of eyes could have been increased or the images broken up into sets of different compression levels and assessed with breaks of 1 week or 1 month. However, it is doubtful whether the effect of memory of images on this type of study can be completely overcome.
As was expected, there was some variation in the terminology that assessors used to describe an abnormality. Guidelines for this could be provided in future studies.
The assessment of overall image quality by the ophthalmologists was also straightforward. They recognized immediately that the 14-KB JPEG compressed images were of poor quality, because the blocking effects were very obvious. However, comments were made anyway that these images would still be suitable if assessing parameters such as optic-disc cupping.
The plots of the RMS error reveal a variation in the effect of image compression on each of the color channels, and a different relative effect on the color channels, depending on the compression technique used. The blue channel after JPEG compression stands out a little in Figure 3 as having higher RMS error. When the images are displayed in their three different channels it can be seen that there is little information in the blue channel and that it contributes little to the full color image; it is quite dark, and varies little in gray level. More variation is seen in the green channel, with most in the red channel (which is also the brightest as would be expected). This is due to the spectral characteristics of the eye. However, it can be considered that loss of information in the blue channel is not as important as it may be to the other channels.
That RMS error is less after Wavelet image compression than that with an image of equivalent size by JPEG compression is due to JPEG compression operating on small blocks of the image, whereas Wavelet compression works on the whole image. Therefore there is some level of averaging in a 8 x 8 block by JPEG compression, which is independent of the neighboring 8 x 8 blocks; at higher levels of compression this manifests itself as the blocking artifact. The differing relative effects on the channels are probably attributable to the methods used for compression. For JPEG compression, the image was compressed as a whole, whereas for Wavelet compression each color channel was compressed separately.
The computational speed of each technique should also be noted. JPEG compression requires very little computational time (a second or less); Wavelet compression, however, requires the image to be analyzed as a whole, which is more memory and computational intensive, and takes in the order of 60 seconds per image. Decompression of compressed images has the same relative time difference. However, these time differentials will decrease with improved software, and an increased computer processing speed.
This last factor can be crucial in determining a choice between selecting JPEG or Wavelet compression. We have shown that a high degree of image compression can be achieved with both methods. If one were to use JPEG to compress a 1.5-MB image to 29 KB, transmission speed is reduced from 25 minutes to 30 seconds. A major attraction of JPEG compression is its popularity with computer systems, software, and users. Most imaging software can read it, and it has wide use on the Internet.
Where computational time was not an issue, a Wavelet compression ratio of 1:68 can reduce this time to 21 seconds. If one considers that in this study we assessed images with subtle abnormalities, it could be argued that if one were screening for more serious, and more obvious, abnormalities, then Wavelet compression down to 15-KB image size (1:100) would still provide good quality images. Transmission time in this case would be 15 seconds. Even then, in cases in which there is some doubt, the ophthalmologist can request a higher resolution image, with the original image being retained uncompressed at the remote site.
Compression algorithms are still evolving. The JPEG committee is working on a new standard named JPEG 2000, which includes a Wavelet algorithm. Yang and Mitra22 use a vector quantization technique to encode radiographs. Although this study has concentrated on the compression of still images, moving images (e.g., from an ophthalmoscope) also play an important role in ophthalmic diagnosis23 and, therefore, compression of video images, for example using MPEG or Apple QuickTime, must also be addressed. Furthermore, because this study was based on scanned 35-mm slides, our results may not apply to images acquired directly from digital retinal cameras and other ophthalmic devices with video attachments. Compression and readability of these images should also be investigated.
|
Conclusions |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionConclusionsReferences |
|---|
|
Acknowledgements |
|---|
|
Footnotes |
|---|
Commercial relationships policy: N.
Corresponding author: Robert H. Eikelboom, Lions Eye Institute, Centre for Ophthalmology and Visual Science, The University of Western Australia, 2 Verdurn Street, Nedlands WA 6009, Australia. robeik{at}cyllene.uwa.edu.au
|
References |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionConclusionsReferences |
|---|
This article has been cited by other articles:
|
|
 |
|
  L Koenig, E Parks, M Analoui, and G Eckert The impact of image compression on diagnostic quality of digital images for detection of chemically-induced periapical lesions Dentomaxillofac. Radiol., January 1, 2004; 33(1): 37 - 43. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 F. Gomez-Ulla, M. I. Fernandez, F. Gonzalez, P. Rey, M. Rodriguez, M. J. Rodriguez-Cid, F. F. Casanueva, M. A. Tome, J. Garcia-Tobio, and F. Gude Response to Merin and Reeves Diabetes Care, March 1, 2003; 26(3): 966 - 967. [Full Text] [PDF]
|
|
|
|
|
|
G. P. Leese, A. Ellingford, A. D. Morris, J. D. Ellis, and S. Cunningham Screening Using Compressed Digital Retinal Images Successfully Identifies Retinopathy Diabetes Care, January 1, 2003; 26(1): 247 - 247. [Full Text] [PDF]
|
|
|
|
|
|
F. Gomez-Ulla, M. I. Fernandez, F. Gonzalez, P. Rey, M. Rodriguez, M. J. Rodriguez-Cid, F. F. Casanueva, M. A. Tome, J. Garcia-Tobio, and F. Gude Digital Retinal Images and Teleophthalmology for Detecting and Grading Diabetic Retinopathy Diabetes Care, August 1, 2002; 25(8): 1384 - 1389. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
