Is digital image compression acceptable within diabetic
retinopathy screening?
Digital
Short
Report
report
image
compression
and DR screening? A. Basu et al.
Oxford,
Diabetic
DME
Blackwell
1464-5491
20
UK
Medicine
Publishing
Science
Ltd,Ltd.
2003
A. Basu, A. D. Kamal, W. Illahi*, M. Khan*, P. Stavrou* and R. E. J. Ryder
Abstract
City Hospital and *Midland Eye Centre,
Birmingham, UK
Accepted 21 March 2003
Aims The National Screening Committee (NSC), whilst recommending the use
of digital mydriatic retinal photography for diabetic retinopathy screening, has
not yet accepted the use of digitally compressed images for grading. By greatly
reducing the file size, however, compression of images is invaluable for storage
and for its rapid transmission across computer networks. We undertook a
study to compare the different levels of JPEG compression with the original
bit-mapped image to determine whether there was any loss of clinical detail
following compression.
Methods Three hundred and thirty images were analysed in this study. These
images had been captured from 66 eyes consecutively photographed in a diabetic retinopathy screening programme, using a Sony DXC-950 P 3CCD colour
video camera mounted on a Canon CR6-45NMf fundus camera. Single 45°
macula-centred images were taken from each eye. The images were compressed
using the JPEG algorithm within Adobe Photoshop (version 4.0) and then displayed with a Sony Trinitron colour monitor. Four different levels of compression were used, JPEG-1, JPEG-2, JPEG-3, JPEG-4, and an objective analysis was
undertaken using ‘lesion counts’. The compressed images were assessed separately and blindly and the results compared with their original BMP images.
Results Eight BMP images could not be evaluated (five right eye and three left
eye). A total of 290 images were therefore used in the final evaluation. All the
JPEG-1 images with file sizes between 16 and 24 kb were found to be ‘pixelated’, while the JPEG-4 images (66–107 kb) appeared similar to the original BMP
(1.3 Mb) images. Both JPEG-2 and JPEG-3 images had significantly lower
counted lesions than the BMP images.
Conclusions From our findings we can conclude that only some degree of image
compression (compression ratios of 1 : 20 to 1 : 12) with file sizes of 66–107 kb
is permissible using JPEG format, whereas the images obtained after higher
compression ratios may not be suitable for diabetic retinopathy screening.
Diabet. Med. 20, 766–771 (2003)
Keywords
digital imaging, retinal screening, image compression, JPEG
Abbreviations CCD, charge-coupled device; JPEG, Joint Photographic Experts
Group
Introduction
Correspondence to: Dr A. Basu, Clinical Research Fellow and Specialist
Registrar in Diabetes, Department of Diabetes, Endocrinology and Metabolism,
City Hospital, Dudley Road, Birmingham B18 7QH, UK.
E-mail: [email protected]
766
The National Screening Committee (NSC) recommends the
use of digital mydriatic retinal photography for diabetic retinopathy screening. It does not, however, recommend the use of
digitally compressed images for grading [1]. Digital retinal
© 2003 Diabetes UK. Diabetic Medicine, 20, 766–771
Short report
767
photography has distinct advantages over conventional 35mm photography despite a lower resolution of the acquired
image, and has proved to be an effective screening modality in
patients with established or partially treated diabetic retinopathy [2]. It allows instantaneous viewing of the retinal images,
and provides an inexpensive way of printing hard copies of the
images [3] that are often superior to Polaroid prints [4]. Image
manipulation using a variety of image editing software may be
used to improve lesion detection during the screening process
[5]. More recently, computer-assisted image analysis is being
used for the development of computer software that may allow
automated retinopathy grading in the future [6– 9].
One of the models of diabetic retinopathy screening process
which arises from the NSC recommendations envisages images
captured out in the field being transmitted electronically to
central grading centres for evaluation [1]. One of the limitations of such transmission over networks is the significant time
delay involved in transmitting large files, which becomes
increasingly problematic as higher resolution images are generated by the digital cameras. High-resolution images generate
large file sizes (as more data per pixel get stored) that slow
transmission and also occupy significant computer memory
space during archiving. One of the solutions to this problem is
to allow some degree of image compression that will maintain
as much clinically relevant detail as the parent image. Among
a variety of compression techniques available, the JPEG compression format is often used for graphical data as the loss of
data is in the form of colour information, which may not be
of clinical significance [10]. We therefore undertook a study
to ascertain whether compressed digital retinal images lose
clinically relevant details when compared with the parent
uncompressed image, i.e. do retinal details get altered when
images are digitally compressed?
This allowed uniformity in the image compression ratios for all
the images. This resulted in five groups of images: BMP, JPEG-1,
JPEG-2, JPEG-3, JPEG-4; BMP represented the original raw image
while JPEG-1, JPEG-2, JPEG-3, and JPEG-4 were image sets that
had maximum, high, medium and low levels of compression,
respectively. The four JPEG groups were evaluated separately
from their bitmaps and these had been blinded to the grader (A.B.).
A lesion count was performed under the following categories: microaneurysms (MA), haemorrhages (HGES), ‘cotton
wool spots’ (CWS), and hard exudates (HE) for each of the images. A strategy using lesion counts to differentiate the images
compared with conventional grading of the images was used as
it was assumed that evaluation of a compressed digital image
from its parent would involve subtle changes in character of the
lesions rather than a change in the overall grade. For example,
an image with four microaneurysms only, correctly classified as
non-proliferative retinopathy, would retain the same grade
even if only three out of the four microaneurysms were visible
after compression. This is extremely important, as it implies a
loss of clinically relevant information and must be taken into
consideration in any study that attempts to evaluate digitally
compressed images.
The lesion counts obtained in each of the JPEG categories
were compared separately with their original BMP images. For
each of these categories, up to six lesions were counted; any
lesion with counts ≥ 6 were labelled as ‘≥ 6’. This was done as we
realized that accurate counting of microaneurysms and haemorrhages is not possible when they appear in clusters. An ‘agreement’ between a pair of images was defined only when identical
lesion counts in each of the above categories (MA, HGES,
CWS, HE) were found. Any discrepancy in this count between
the JPEG image and its parent BMP was defined as a ‘disagreement’ in the retinopathy. Lesion counts to determine difference
between colour photographs and fluorescein angiograms in
detecting diabetic retinopathy have been used before [11]. We
have used a similar approach.
Patients and methods
Results
Sixty-six right and left eye images were captured from 33 consecutive patients using a Sony DXC-950 P 3CCD colour video
camera mounted on a Canon CR6-45NMf camera. The patients were part of the routine attendance at diabetes retinal
screening clinics at City Hospital, Birmingham, UK. The median
age was 57 years (range 27–81 years). Most had Type 2 diabetes (n = 31). They were of mixed ethnicity: 14 Caucasians, 11
Asians and eight Afro-Caribbeans. Informed verbal consent
was obtained from all patients. All patients had checks of visual
acuity and instillation of mydriatic drops (1% Tropicamide)
prior to photography. Single 45° macula-centred images were
taken from each eye. The images were captured using ‘EyeCap’
software (Orion Imaging Ltd, Cumbran, UK) and a Pentium III PC
and saved in Windows Bitmap (BMP) format on CD-ROM. For
the purposes of this study the images were displayed using a
Sony colour monitor [display resolution 1024 × 768 pixels and
24-bit (True Colour) colour display] and viewed using Adobe
Photoshop (version 4.0) image editing software. This software
was used to compress each individual image into four levels of
JPEG compression (low, medium, high, maximum) using the
JPEG compression algorithm specified within the software.
Of the 66 BMP images, eight could not be graded (two BMP
images had been incorrectly saved and six images had significant cataracts rendering them ungradable). A total of 290
images were therefore evaluated. The retinopathy grades for
these 66 BMP images are shown in Table 1.
© 2003 Diabetes UK. Diabetic Medicine, 20, 766–771
Objective assessment of images using lesion counts
All the JPEG-1 images were found to be ‘pixelated’, which did
not allow clear definition of lesion or vessel margins. These
Table 1 Retinopathy grades (n = 66)
Retinopathy grade
Number of images
No diabetic retinopathy
Non-proliferative retinopathy
Proliferative retinopathy
Not gradable
26
26
6
8
768
Digital image compression and DR screening? • A. Basu et al.
Agreement
Disagreement
Total number of pixelated images
JPEG-1
16 –24 kb
JPEG-2
24 –33 kb
JPEG-3
37– 61 kb
JPEG-4
66–107 kb
0 (0)
0 (0)
58 (100%)
36 (62%)
9 (16%)
13 (22%)
52 (90%)
6 (10%)
0 (0)
58 (100%)
0 (0)
0 (0)
Table 2 Comparison of compressed images
with Bitmap images (1270 kb)
Numbers within cells represent the number of images.
Table 3
a. Frequency of BMP images among the different lesion types
Lesion count categories
Lesion types
≤2
3–5
≥6
Microaneurysms (MA)
Haemorrhages (HGES)
Cotton wool spots (CWS)
Hard exudates (HE)
8
11
1
2
5
4
1
0
11
8
0
4
b. Reasons for disagreements
Reasons for disagreement—lesion counts JPEG/BMP
Number
of images
JPEG-2
JPEG-3
MA
HGES
Right eye
Left eye
Total
2
7
9/58
MA 3/≥ 6; MA 4/5
MA 3/6; MA 1/3; MA 2/3; MA 2/5; MA 2/3
HGES 3/4; HGES 2/3
Right eye
Left eye
Total
1
5
6/58
MA 4/≥ 6
MA 4/6; MA 2/3; MA 5/6; MA 4/5
HGES 1/2
Numbers within cells represent the number of images within each category.
images had files sizes of 16–24 kb and were as a result of 1 : 52
to 1 : 79 compression. On the other hand, none of the JPEG-3
and JPEG-4 images was found to have been ‘pixelated’.
Difference in images (disagreement based on lesion counts
+ total number of pixelated images) was noticed in 22/58
JPEG-2 images and 6/58 JPEG-3 images (Table 2). The data
are expressed as a proportion of images in agreement or disagreement/total number of images. As this is not a k × k grading
system but rather a 1 × k grading system (‘1’ representing the
BMP and ‘k’ the different categories of JPEG compression), the
κ statistic cannot be used. The ‘agreement’ between the images
here is based on having identical lesion counts in different categories (MA, HGES, CWS, HE) between the pair of images (a
‘BMP’ and a ‘JPEG’), where the BMP image is the ‘reference
standard’. Microaneurysms and haemorrhages were the predominant lesions encountered in our data set of images
(Table 3a). The disagreement noted was predominantly due to
reduced counting of microaneurysms in the JPEG-2 and JPEG3 images compared with their original BMP image; in JPEG-2
images 7/9 images and in JPEG-3 5/6 were as a result of a
reduced counting of microaneurysms (Table 3b). This is not
surprising, as small red lesions such as microaneurysms and
small haemorrhages are normally difficult to detect against a
red background, and it is still more likely to be the case when
the images are compressed. Although the total number of
images with hard exudates and cotton wool spots was less than
those with microaneurysms and haemorrhages, little difference was noticeable in these lesions after compression. This
may be because these lesions, being white, are better contrasted against a red retinal background and therefore more
easily detectable. JPEG-1 images of course could not be compared, as they were all universally ‘pixelated’.
Side-by-side comparison of the original BMP images and the
JPEG-4 images revealed that they were identical as far as could
be assessed by the human eye. This was confirmed after careful
scrutiny of the images by several of the authors (A.B., R.E.J.R.,
W.I., A.D.K.) ( Figs 1 and 2).
Discussion
Digital images can be stored in a variety of compressed file formats. Broadly these are categorized as ‘lossy’ compression or
‘lossless’ compression formats. With ‘lossy’ compression techniques, e.g. JPEG, it is impossible to recreate the original image
from the compressed image due to loss of mathematical information. In some instances the compression algorithm used
© 2003 Diabetes UK. Diabetic Medicine, 20, 766–771
Short report
Figure 1 Image compression of a single retinal image.
© 2003 Diabetes UK. Diabetic Medicine, 20, 766–771
769
770
Digital image compression and DR screening? • A. Basu et al.
Figure 2 Corresponding areas from the images cropped to focus on the subtle changes. Finer details of vessels and lesions, especially around their edges,
are lost with progressive compression, whilst the grade of retinopathy remains unaltered.
may be completely reversed so that the decompressed image is
‘digitally’ identical to the original—this is described as ‘lossless’
compression, e.g. TIFF. Despite being a ‘lossy’ compression
technique, JPEG compression has a number of advantages. It
maintains brightness and contrast information (to which the
human eye is very sensitive) at the expense of some colour definition, and is therefore useful for images from film or video
[12]. Second, the user can conveniently define the amount of
compression necessary by adjusting certain compression
parameters. Although this allows the user to trade off file size
against output image quality, this difference in quality may not
be noticeable to the human eye up to a certain level of compression. For this reason the JPEG algorithm is the one most commonly used for images transmitted over the internet.
A number of studies have looked at digital compression of
medical images with relative preservation of image quality
following compression [13 –15]. Digital image compression
(JPEG) of retinal and optic nerve head images for the purposes
of tele-ophthalmology has been studied before on a per-lesion
basis [16,17]. A study [18] attempting to evaluate the effect of
image compression in the context of diabetic retinopathy
screening using digitized 35-mm slides not surprisingly found
that the sensitivity reduced with increasing degrees of JPEG
compression. However, one of the limitations of this study was
in the use of digitized images whose quality is dependent on the
scanning technique and the scanners used to generate the
slides. It is therefore difficult to draw any definitive conclusions from their results. In the case of digitally acquired images
as are generated by modern digital retinal cameras, the image
quality is specified at the time of capture.
In our study we used digitally acquired images, as would be
expected during routine retinopathy screening. We also
emphasize that while it is important to detect changes in individual lesions following compression, viewing these changes
within the context of the entire retinal image is necessary
rather than the component parts in isolation as has been done
before [17]. In other words, whilst it may be possible to detect
hard exudates clearly in one image, it may not be possible to
detect a small microaneurysm in the same image, and this
is undoubtedly important from the grader’s perspective.
Although the total number of images in our data set was small,
40% (23/58) of the images had a large number (≥ 6) of lesions
per image. However, the numbers of images with hard exudates and cotton wool spots were few, and whilst we feel that
our findings are unlikely to be different even if a larger data set
were used, this could be further confirmed using images with
many such lesions per image. In this study we used a specific
range of compressions and analysed their effects on diabetic
retinopathy lesions. Therefore, whilst we can conclude that
some degree of image compression may be acceptable [such as
the JPEG-4 images with file sizes of 66–107 kb (compression
ratios of 1 : 20 to 1 : 12)], we cannot use our study results to
specify the optimum compression ratio that can be accepted as
a standard across the board. We also recognize that some grading centres may have large computer storage capacities and
have transmission lines that allow a broad bandwith, in which
case such image compression may not be necessary. However,
with the availability of less expensive high-resolution digital
cameras (such as the Canon D-30), large image file sizes will be
inevitable and one may then need to consider carefully some
© 2003 Diabetes UK. Diabetic Medicine, 20, 766–771
Short report
image compression options. The optimum JPEG compression
ratios for these higher resolution images, however, need to be
studied separately.
Our study may open an opportunity for diabetic retinopathy
screening centres to evaluate JPEG compressed images at a local
level, and decide their optimum method for archiving and transmission of such images electronically and could also perhaps
be carefully considered within the national screening guidelines.
Appendix 1: useful definitions [12,19]
Pixel: This is the short for ‘picture elements’; it represents the
smallest point in a graphic image.
Pixelated: The appearance of pixels in a bit-mapped image—
when an image is displayed or printed too large the individual
square pixels are discernible to the naked eye.
Bit-mapped image: The representation of an image as rows
and columns of dots, in the computer’s memory. The value of
each dot (whether it is filled in or not) is stored in one or more
‘bits’ of data. The density of the dots, known as the resolution,
determines how sharply the image is represented. This is often
expressed in dots per inch (dpi) or simply by the number of
rows and columns, such as 640 × 480. These bit-mapped
images can be stored as a BMP (MS Windows environment),
TIFF (tagged image file format) or PCX files.
JPEG image ( Joint Photographic Experts Group): JPEG
(pronounced ‘jay-peg’) is a standardized image compression
mechanism. This compression algorithm is now available in
most image-editing software, and in our study we used Adobe
Photoshop (version 4.0).
Acknowledgements
We wish to acknowledge Mr Tim Marshall, University of
Birmingham, for statistical advice and help in preparation of
this manuscript and the Medical Illustrations Department, City
Hospital, for capturing the images for this study.
References
1 A model for a photographic diabetic retinopathy screening programme:
http://www.diabetic-retinopathy.screening.nhs.uk/photographic-model/
index.html
2 Kerr D, Cavan DA, Jennings B, Dunnington C, Gold D, Crick M. Beyond
retinal screening: digital imaging in the assessment and follow-up of
patients with diabetic retinopathy. Diabet Med 1998; 15: 878–882.
© 2003 Diabetes UK. Diabetic Medicine, 20, 766–771
771
3 Game FL, Kong N, Illahi W, Ryder REJ. Important practical considerations with regard to file-type, file-size, display-mode and ‘hardcopy’ in digital imaging for diabetic retinopathy screening. Diabet
Med 1997; 14: S11–12.
4 Ryder REJ, Kong N, Bates AS, Sim J, Welch J, Kritzinger EE.
Instant electronic imaging are superior to Polaroid at detecting
sight-threatening diabetic retinopathy. Diabet Med 1998; 15: 254–
258.
5 George LD, Lusty J, Owens DR, Ollerton RL. Effect of software
manipulation (Photoshop) of digitised retinal images on the
grading of diabetic retinopathy. Br J Ophthalmol 1999; 83: 911–
913.
6 Gardener CG, Keating D, Williamson TH, Elliot AT. Automated
detection of diabetic retinopathy using an artificial neural network:
a screening tool. Br J Ophthalmol 1996; 80: 940–944.
7 Cree MJ, Olson JA, McHardy KC, Sharp PF, Forrester JV. A fully
automated comparative microaneurysm digital detection system. Eye
1997; 11: 622–628.
8 Sinthanayothin C, Boyce JF, Williamson TH, Cook HL, Mensah E,
Lal S et al. Automated detection of diabetic retinopathy on digital
fundus images. Diabet Med 2002; 19: 105–112.
9 Hipwell JH, Strachan F, Olson JA, McHardy KC, Sharp PF,
Forrester JV. Automated detection of microaneurysms in digital
red-free photographs: a diabetic retinopathy screening tool. Diabet
Med 2000; 17: 588–594.
10 The Joint Photographic Experts Group. http://www.jpeg.org
11 The DCCT Research Group. Color photography vs fluorescein
angiography in the detection of diabetic retinopathy in the Diabetes
Control and Complications Trial. Arch Ophthalmol 1987; 105:
1344–1351.
12 Brinkmann R. The digital representation of visual information.
In The Art and Science of Digital Compositing. Sidcup: Morgan
Kaufmann, 1999.
13 Martin C, Packer S, Lesser M. Perceived quality of digital fundus
images (ARVO Abstract). Invest Ophthalmol Visual Sci 1991; 32
(Abstr. 127): S690.
14 Persons K, Palisson P, Manduca A, Erickson BJ, Savcenko V. An
analytical look at the effects of compression on medical images. J
Digit Imaging 1997; 10: 60–66.
15 Bittorf A, Fartasch M, Schuler G, Diepgen TL. Resolution requirements for digital images in dermatology. J Am Acad Dermatol 1997;
37: 195–198.
16 Yogesan K, Constable IJ, Morgan W, Soebadi DY. International
transmission of tele-ophthalmology images. J Telemedicine Telecare
2000; 6: 41–44.
17 Eikelboom RH, Yogesan K, Barry CJ, Constable IJ, Tay-Kearney M,
Jitskaia L et al. Methods and limits of digital image compression of
retinal images for telemedicine. Invest Ophthalmol Visual Sci 2000;
41: 1916–1924.
18 Newsom RSB, Clover A, Costen MTJ, Sadler J, Newton J, Luff AJ
et al. Effect of digital image compression on screening for diabetic
retinopathy. Br J Ophthalmol 1999; 85: 799–802.
19 Freedman A. The Computer Desktop Encyclopaedia, 2nd edn. New
York: AMACOM Publications, 1999.