JPEG compression of stereoscopic digital images for the diagnosis of diabetic retinopathy via teleophthalmology

JPEG compression of stereoscopic digital images for the diagnosis of diabetic retinopathy via teleophthalmology

JPEG compression of stereoscopic digital images for the diagnosis of diabetic retinopathy via teleophthalmology Chad F. Baker,* MD; Christopher J. Rud...

899KB Sizes 0 Downloads 62 Views

JPEG compression of stereoscopic digital images for the diagnosis of diabetic retinopathy via teleophthalmology Chad F. Baker,* MD; Christopher J. Rudnisky,* MD; Matthew T.S. Tennant,* MD, FRCSC; Paul Sanghera, t MD; Bradley J. Hinz,* MD, FRCSC; Alexander R. De Leon,:j: PhD; Mark D.J. Greve,* MD, FRCSC ABSTRACT • RESUME Background: Canada's vast size and remote rural communities represent a significant hurdle for successful monitoring and evaluation of diabetic retinopathy.Teleophthalmology may provide a solution to overcome this problem. We investigated the application of Joint Photographic Experts Group UPEG) compression to digital retinal images to determine whether JPEG compression could reduce file sizes while maintaining sufficient quality and detail to accurately diagnose diabetic retinopathy. Methods: All 20 patients with type 2 diabetes mellitus assessed at a 1-day teleophthalmology clinic in northern Alberta were enrolled in the study. Following pupil dilation, seven 30° fields of each fundus were digitally photographed at a resolution of 2008 X 3040 pixels and saved in uncompressed tagged image file format (TIFF). The files were compressed approximately 55x and I 13x their original size using JPEG compression. A reviewer in Edmonton randomly viewed all original TIFF images along with the compressed JPEG images in a masked fashion for image quality and for specific diabetic retinal pathology in accordance with Early Treatment Diabetic Retinopathy Study standards. The level of diabetic retinopathy and recommendations for clinical follow-up were also recorded. Exact agreement and weighted K statistics, a measure of reproducibility, were calculated. Results: Exact agreement between the compressed JPEG images and the TIFF images was high (75% to I00%) for all measured variables at both compression levels. Reproducibility was good to excellent at both compression levels for the identification of diabetic retinal abnormalities (K = 0.45-1 ), diagnosis of level of retinopathy (K = 0.73-1) and recommended follow-up (K = 0.64-1 ). Interpretation: The application of JPEG compression at ratios of 55: I and 113: I did not significantly interfere with the identification of specific diabetic retinal pathology, diagnosis of level of retinopathy or recommended follow-up. These results indicate that JPEG compression at ratios as high

From *the Department of Ophthalmology, University of Alberta, Royal Alexandra Hospital, Edmonton, Alta., tthe Department of Ophthalmology, University of Toronto, Toronto, Ont., and *the Department of Mathematics and Statistics, University of Calgary, Calgary, Alta.

Correspondence to: Dr. Matthew T.S. Tennant, Alberta Retina Consultants, Suite 400, ]0924 107 Ave., Edmonton AB TSH OXS; mtennant@ alberta-retina.com This article has been peer-reviewed.

Originally received Feb. 20, 2003 Accepted for publication Aug. 11, 2004

746

Can j Ophthalmol 2004;39:746-54

JPEG compression of retinal images-Baker et al

JPEG compression of retinal images-Baker et al

as I 13: I has the potential to reduce storage requirements without interfering with the accurate and reproducible teleophthalmologic diagnosis of diabetic retinopathy. This pilot project demonstrates the potential for JPEG compression within a digital teleophthalmology viewing system.

·Contexte: La vaste etendue du Canada et l'eloignement des collectivites rurales font grandement obstacle au monitorage eta !'evaluation de Ia retinopathie diabetique. La teleophtalmologie peut presenter une solution a ce probleme. Nous avons cherche a savoir si !'application de Ia technique de compression Joint Photographic Expert Group UPEG) a l'image numerique de Ia retine permettait de reduire Ia taille des images tout en conservant suffisamment de qualite et de details pour diagnostiquer Ia retinite diabetique avec exactitude. Methodes : Les 20 patients atteints de diabete sucre de type 2 evalues durant une journee de clinique de teleophtalmologie du Nord de !'Alberta ont ete inscrits a l'etude.Apres dilatation de Ia pupille, I' on a pris et enregisre en format non compresse TIFF (tagged image file format) des photographies numeriques de sept champs de 30° de chaque fond d'ceil a une resolution de 2 008 X 3 040 pixels. Les fiches ont ensuite ete compressees a des taux d'environ 55: I et 113: I en format JPEG. Un examinateur a Edmonton a examine au hasard et de fa~on masquee les images originales en TIFF et celles qui avaient ete compressees en JPEG pour en verifier Ia qualite et pour y deceler de fa~on specifique une pathologie diabetique de Ia retine en regard des normes de Ia Early Treatment Diabetic Retinopathy Study. On a aussi pris en note le niveau de Ia retinopathie diabetique et les recommandations quant au suivi clinique. On a calcule Ia concordance et les statistiques ponderees 11:, mesure de reproductibilite. Resultats: L'exactitude de Ia concordance entre les images compressees JPEG et les images TIGG s'est averee elevee (75% a I00 %) pour toutes les variables mesurees aux deux niveaux de compression. La reproductibilite etait bonne a excellente aux deux niveaux pour !'identification des pathologies diabetiques de Ia retine (11: = 0,45-1), le diagnostic du niveau de Ia retinopathie (11: = 0,73-1) et le suivi recommande (11: = 0,64-1).

Interpretation : L'application de Ia compression JPEG a des taux de 55: I et 113: I n'a pas perturbe de fa~on significative !'identification de Ia pathologie diabetique particuliere de Ia retine,le diagnostic du niveau de Ia retinopathie ni du suivi recommande. Les resultats indiquent que Ia compression JPEG a des degres aussi eleves que Ill: I peut reduire les besoins de stockage sans compromettre !'exactitude et Ia reproductibilite du diagnostic teleophtalmologique de Ia retinopathie diabetique. Ce projet pilote demontre le potentiel de Ia compression JPEG dans un systeme de visionnement numerique en teleophtalmologie.

B

ecause diabetic retinopathy is a leading cause of blindness in North America, 1 and because Canada is a country with a low population density and enormous geographic size, there has been considerable interest in discovering more effective methods to identify diabetic retinopathy. In addition, aboriginal Canadians, with a prevalence of diabetes mellitus reported to be 2.3 2 to 6.7 3 times that of Canadians of European ancestry, are in large part concentrated in rural areas at a distance from ophthalmic care. 4

Traditional methods, whereby a general ophthalmologist or retinal specialist travels to a remote location to evaluate patients with diabetes, are inefficient given the high volume of patients these doctors are required to care for within urban centres. Screening strategies for diabetic retinopathy have been under investigation for many years. We propose an alternative to the notion of "screening" and offer instead the opportunity to perform distance evaluation using stereoscopic digital fundus photography.

CAN J OPHTHALMOL-VOL. 39, NO.7, 2004

747

JPEG compression of retinal images-Baker et al Fort Vermilion, Alta., located 640 km north of Edmonton, is the site of our teleophthalmology pilot project to improve diabetic eye care for rural Canadians. In Canada, it is recommended that all patients with type 2 diabetes be evaluated at least annually following diagnosis. 5 Owing to geographic and cultural obstacles, this standard is often not achieved in rural areas. We are attempting to increase rural accessibility to specialist eye care using a novel stereoscopic digital imaging system. Advances in high-resolution digital imaging in recent years have allowed this technology to become a suitable replacement for slide film when photographing retinal disease. 6-9 However, the use of highresolution digital images consumes significant computer storage space. In addition, these bulky files dramatically increase transmission and processing times and tax all but the most sophisticated systems. To combat the problems produced by large image file sizes, digital image compression needs to be explored. We carried out a pilot study to investigate the application of Joint Photographic Experts Group (JPEG) compression to digital photographic data obtained during a single day of clinic. METHODS

Patients in a teleophthalmology clinic held in Fort Vermilion for 1 day in July 2001 were asked to participate in a pilot study examining the efficacy of digital image compression. The study was designed as a preliminary project with a relatively small number of participants rather than as a large, definitive study. Inclusion criteria for the study were a history of diabetes mellitus diagnosed by a physician and willingness to participate in the study. Informed consent was obtained from all patients. All 20 patients assessed at the teleophthalmology clinic were included. Before retinal photography, visual acuity and intraocular pressure were measured by a trained technician. Patients' eyes were then dilated with a combination of 2.5% phenylephrine hydrochloride and 1.0% tropicamide (Diophenyl-T, Dioptic Laboratories, Markham, Ont.). Seven 30° fundus photographs were captured by a certified ophthalmic photographer using a fundus camera (Zeiss FF450, Carl Zeiss Corporation, Jena, Germany) 10 with a high-resolution digital camera back (DCS 560, Eastman Kodak, Rochester, NY [resolution 2008 x 3040 pixels]). Stereoscopic images were captured of

7 48

CAN J OPHTHALMOL-VOL. 39, NO. 7, 2004

the disc and macula (fields 1 and 2) through a cornealinduced parallax, 11 and nonstereoscopic images were obtained of fields 3 through 7. All digital images were transferred to a laptop computer for storage. Before analysis, all images were assessed for optimal exposure and, if necessary, were adjusted to optimize viewing (Twain software, Eastman Kodak). The original uncompressed tagged image file format (TIFF) files were then compressed (Adobe Photoshop, version 5.5, Adobe Systems Incorporated, San Jose, Calif.) to approximately 55 times (quality factor 7) and 113 times (quality factor 4) their original file size into JPEG format. To facilitate analysis in a masked and unbiased manner, we used a random-number generator to rename and randomize the three data sets. Image sets consisting of seven fields of both the right and left fundi were displayed in a masked fashion on a 56-cm high-definition video monitor (model P225, ViewSonic Corporation, Walnut, Calif.) at a resolution of 1024 x 768 pixels in 32-bit colour using a threedimension-enabled OpenGL video card (Nvidia Quadro 4 980 XGL, Nvidia Corporation, Santa Clara, Calif.). The digital photographs were scaled automatically to display the entire image on the screen. In addition, the reviewer (M.T.S.T.) had the ability to magnify the images at his discretion. Proprietary stereoviewing software (Twinviewer, University of Alberta, Edmonton) and liquid crystal shutter goggles (Stereographics Corp., San Rafael, Calif.) were used to view the stereoscopic digital photographs of fields 1 and 2. Each individual digital image was analysed for the presence or absence of microaneurysms, intraretinal hemorrhage, hard exudates, cotton-wool spots, intraretinal microvascular abnormalities, venous beading, macular edema, clinically significant macular edema, neovascularization of the disc, neovascularization elsewhere in the retina, vitreous or preretinal hemorrhage, and severe neovascularization. Image quality was also recorded for each eye and was graded "acceptable" if more retinal detail was visible than in Early Treatment Diabetic Retinopathy Study (ETDRS) standard photo 14. The reviewer graded the level of diabetic retinopathy found in each standard seven-field image set using the ETDRS extension of the modified Airlie House classification. 10 Diabetic retinopathy was recorded in accordance with ETDRS standards as no diabetic retinopathy (level 10), mild nonproliferative diabetic retinopathy (NPDR) (level 35), moderate NPDR (level 43), severe NPDR (level 53A-D) or very severe NPDR (level 53E). Pro-

JPEG compression of retinal images-Baker et al

liferative diabetic retinopathy (PDR) was classified as early PDR (level61 or 65) or high-risk PDR (level 71 or 75). 12 The diagnoses of microaneurysms only (level 20) and mild NPDR are based on the observation of occasionally indiscriminate lesions, and they typically result in the assignment of similar treatment plans. Therefore, we decided to merge these two diagnostic categories for the purposes of this study. We used standard ETDRS diabetic retinal photographic slides, scanned at high resolution and digitally displayed with the same hardware as used for the patient photographs, to aid in image analysis. After grading the extent of diabetic retinopathy present in each randomized set of images, the reviewer recorded his recommendations for patient follow-up. We calculated the exact agreement between the grading of the original TIFF images and that of the compressed JPEG files. Then, to establish the degree of reproducibility, we calculated a weighted version of the intraclass correlation coefficient K 13 •14 for the 20 complete image sets of left and right eyes at each compression level. For data analysis we used an approach that recognizes the paired nature of the information obtained from the images of the left and right eyes. As a result, we avoided the need to carry out separate analysis of the left and right eyes for each patient, as this would not take into account the possible correlation of pathological findings between a patient's eyes. 14•15 Existing formulae for calculating standard error applies to this approach for binocular data 16 such that there was no need to adjust the standard errors. 17 A full weight of 1 was given to pairs of eyes where the original TIFF files and the compressed JPEG images showed the same retinal abnormality, level of diabetic retinopathy or follow-up recommendations for both eyes, and a weight of 0.5 to those where there was agreement for only one eye. K values of 0.4 or less indicated marginal reproducibility, values of 0.41 to 0.75 indicated good reproducibility, and values greater than 0.75 indicated excellent reproducibility. 18 We constructed approximate 95% confidence intervals for the K values calculated for each of the 13 diagnostic categories. In addition, we carried out tests of the hypothesis K = 0 at a level of significance of 5%. RESULTS

The mean age of the 20 patients was 63.7 years (range 29-83 years, median 56 years). Eleven of the patients were men, and nine were women. The time

since diagnosis of diabetes ranged from 1 to 20 years (mean 11.6 years, median 10.5 years). The primary method of glucose control was oral hypoglycemic agents in 10 patients, insulin in 9 and diet in 1. All patients enrolled were included in the final analysis. Because digital stereoimages were captured of fields 1 and 2 in each eye as well as monoscopic images of fields 3-7, there were a total of 18 images for each patient. Image quality was acceptable in all cases. The extent of exposure compensation ranged from -0.9 for overexposed images to +1.2 for underexposed images. The original 17.4-MB TIFF images required a total of 6280 MB of digital memory for the 360 images viewed. JPEG compression utilities in Adobe Photoshop, version 5.5, reduced the file size according to quality factor settings rather than by specific ratios. The extent of compression at a set factor depends on the amount of photographic detail and colour variation in each image. 19 As a result, the compression of each image by a specific factor did not produce files of uniform size. Reducing a TIFF image to JPEG quality factor 7 produced a mean file size of 0.318 MB (range 0.259-0.375 MB) and a total memory requirement of 115MB (a decrease of about 55x). Reducing to JPEG quality factor 4 produced a mean file size of 0.154 MB (range 0.135-0.188) and a total memory requirement of 55.4MB (a decrease of about 113x). The numbers of eyes with each retinal abnormality and the level of diabetic retinopathy, as assessed by examination of the uncompressed TIFF image sets, are shown in Tables 1 and 2. The recommendations for patient follow-up based on these images are given in Table 3. There was a high level of exact agreement for both levels of image compression (Table 1). For the images compressed 55x, the rate of exact agreement ranged from 75% for macular edema to 100% for neovascularization of the disc and neovascularization elsewhere. For the diagnosis of level of diabetic retinopathy, the rate of exact agreement ranged from 90% for mild NPDR to 100% for moderate NPDR, severe NPDR, early PDR and high-risk PDR (Table 2). For patient follow-up recommendations, rates of exact agreement ranged from 80% (1 year) to 95% (6 months and immediate referral) (Table 3). Similar rates of exact agreement were found for the images compressed 113x. Good to excellent correlation was obtained between

CAN JOPHTHALMOL-VOL. 39, NO.7, 2004

749

JPEG compression of retinal images-Baker et al

Table !-Detection of diabetic retinal abnormalities on JPEG images compressed SSx and lllx as compared to uncompressed TIFF images I 13x compression

SSx compression

Pathology Microaneurysms lntraretinal hemorrhage Microaneurysms/ intraretinal hemorrhage Hard exudates Cotton-wool spots lntraretinal microvascular abnormalities Venous beading Macular edema Clinically significant macular edema Neovascularization other than of disc Neovascularization of disc Vitreous or preretinal hemorrhage Severe neovascularization

K

Standard error

Exact agreement rate,%

No. of eyes

Exact agreement rate,%

22 IS

85 95

0.85 0.95

0.078 0.053

23 12 9

90 90 80

0.90 0.88 0.61

6 0 9

85 100 75

5

K

Standard error

90 95

0.90 0.95

0.067 0.051

0.067 0.087 0.150

85 90 80

0.85 0.82 0.61

0.078 0.130 0.150

0.46 Undefined 0.63

0.200 Undefined 0.130

90 100 75

0.77 Undefined 0.60

0.140 Undefined 0.160

90

0.77

0.130

95

0.89

0.110

2

100

1.00

0.00

95

0.66

0.012

I

100

1.00

0.230

100

1.00

0.00

0 I

100 95

Undefined 0.66

Undefined 0.012

100 95

Undefined 0.66

Undefined 0.012

Note:JPEG =Joint Photographic Experts Group;TIFF =tagged image file format.

the compressed and original TIFF images for the identification of all observed diabetic retinal abnormalities, at both levels of compression (Table 1). Because there were no patients in this sample with venous beading or vitreous or preretinal hemorrhage, the K statistic for these abnormalities was calculated to be undefined. However, there were no false-positive results at either compression level for these pathologies. Good correlation was obtained at both levels of compression for the identification of intraretinal microvascular abnormalities, cotton-wool spots, macular edema and severe neovascularization (K = 0.46-0.66). However, there was a slight variation in the K statistic for intraretinal microvascular abnormalities between 55x compression (K = 0.46) and 113x compression (K = 0.78). This finding suggests that the more compressed images were better correlated to the uncompressed TIFF images. We performed a z test on this variation, but it was not significant at the 5% level (z =-1.328, p =0.092). Because we had grouped the diagnoses of microaneurysms and

750

CAN J OPHTHALMOL-VOL. 39, NO.7, 2004

mild NPDR, we examined what would happen if we also grouped microaneurysms and intraretinal hemorrhage into one category (Table 1). A positive response (presence of microaneurysms or intraretinal hemorrhage or both) was recorded for this category if one or both of the abnormalities were identified in a photograph. The rates of exact agreement (90% for the image compressed 55x and 85% for that compressed 113x) and K values indicated excellent correlation with the TIFF images, which implies that grouping these abnormalities together does not adversely affect their identification on compressed images. Excellent correlation was obtained for grading of the level of diabetic retinopathy in the images compressed 55x (K =0.84-1) (Table 2). Grading of those compressed 113x was nearly as accurate, with excellent correlation for all levels of retinopathy except mild NPDR (K = 0.73). There were no patients with very severe NPDR in the sample. Both levels of compression correctly demonstrated the absence of very severe NPDR. This resulted in a rate of

JPEG compression of retinal images-Baker et al

Table 2-Grading of ETDRS level of diabetic retinopathy on JPEG images compressed 55x and 113x as compared to uncompressed TIFF images 55x compression Level of diabetic retinopathy No diabetic retinopathy Mild NPDR Moderate NPDR Severe NPDR Very severe NPDR Early PDR High-risk PDR Note: ETDRS pathy.

No. of eyes

16 9 3 I

0 3 I

Exact agreement rate,%

95 90 100 100 100 100 100

I 13x compression Exact agreement rate,%

K

Standard error

0.90 0.84 1.0 1.0

0.099 0.11 0 0

Undefined

Undefined

1.0 1.0

0 0

85 85 100 100 100 100 100

K

Standard error

0.80 0.73 1.0 1.0

0.11 0.15 0 0

Undefined

Undefined

1.0 1.0

0 0

= Early Treatment Diabetic Retinopathy Study; NPDR =nonproliferative diabetic retinopathy; PDR =proliferative diabetic retino-

Table 3-Recommendati ons for patient follow-up based on JPEG images compressed 55x and 113x as compared to uncompressed TIFF images 55x compression Recommended follow-up I yr

6 mo 3 mo Immediate referral

No. of eyes

Exact agreement rate,%

27 2 6 5

I I 3x compression

K

Standard error

Exact agreement rate,%

80 95 85

0.75 0.64 0.67

0.11 0.33 0.15

95

0.87

0.10

exact agreement of 100% with an undefined K value. Correlation for recommendations for patient follow-up was excellent at both compression levels for yearly follow-up and for immediate referral (Table 3). For follow-up at 3-month intervals and at 6-month intervals, good correlation was observed with both compression levels. INTERPRETATION

Methods of digital image compression may be divided into two categories: lossless and lossy. Lossless compression capitalizes on redundancy within an image to reduce file size without any loss of detail and results in perfect image reconstruction. However, lossless methods are capable of achieving compression levels of only 2:1 to 3:1 for medical applica-

K

Standard error

85 95 85

0.77 0.65 0.67

0.13 0.33 0.15

100

1.0

0.0

tions. 20•21 In contrast, lossy image compression is able to achieve any level of compression by sacrificing minute details present in the original images. As the amount of compression increases, the extent to which the compressed image differs from the original also increases. However, studies of radiologic applications have shown that lossy compression at ratios as high as 80: 1 does not significantly alter the overall diagnostic accuracy of compressed images. 22- 25 JPEG compression is a lossy image compression technique that is designed to take advantage of characteristics of human visual physiology. Specifically, this compression method exploits the fact that small changes in colour are perceived less accurately than slight changes in brightness. In addition, the human perceptual system is able to "fill in" proper detail when it is provided with comparatively crude

CAN I OPHTHALMOL-VOL. 39. NO.7. 2004

751

JPEG compression of retinal images-Baker et al information. 26 Consequently, JPEG compression may be used to generate significantly compressed images with virtually imperceptible differences from the original. Details of the complex mathematical processes underlying JPEG compression can be found in the article by Smith and colleagues. 19 Several of the advantages of a digital stereo acquisition and viewing system have been previously described. 9 •14 Our experience with teleophthalmology to date suggests that, on gross inspection, JPEG images are no different from the uncompressed images retrieved directly from our digital cameras. Anagnoste and associates 27 reported that most of the 168 retinal specialists attending a national meeting could not differentiate 28x JPEG compressed fundus images from the original photographs. Lee and coworkers 25 also found that experienced observers could not consistently or accurately distinguish retinal photographs that had been scanned at 750 dots per inch and then compressed by a ratio of 30: 1 from TIFF images. In a study in diagnostic radiology, original digital chest x-ray films could not be readily identified when compared with JPEG digital images that had been compressed up to 30x. 19 These reports together with our own observations highlight the potential of JPEG compression to reduce computer memory requirements without interfering with accurate diagnosis of diabetic retinopathy. The results of this pilot study suggest that JPEG images at compression levels of 55x and 113x may have near-diagnostic equivalency to uncompressed TIFF images. In comparison with the original TIFF digital images, we found a rate of exact agreement of 75% or better, along with good to excellent reproducibility (K > 0.4) for the detection of most types of diabetic retinal pathology. Unfortunately, the absence of venous beading and vitreous or preretinal hemorrhage in our sample precludes specific comment on the effects of JPEG compression on their identification. None the less, our results suggest that JPEG compression of up to 113:1 should not significantly interfere with the identification of most diabetic retinal abnormalities. We found excellent correlation between the JPEG compressed and TIFF images for all levels of retinopathy and both levels of compression except mild NPDR at 113x compression and very severe NPDR at both compression levels. The diagnosis of mild NPDR is based on the observation of relatively subtle changes, such as microaneurysms and small intraretinal dot

752

CAN J OPHTHALMOL-VOL. 39, NO.7, 2004

hemorrhages. It is possible that these less distinct abnormalities are more susceptible to the distortion produced by JPEG image compression, particularly because the reproducibility decreased at a higher level of compression. However, good reproducibility (K = 0. 725) of the diagnosis of mild NPDR was achieved at 113x compression. Since there were no patients with very severe NPDR in our study, we cannot draw conclusions regarding the diagnosis of this level of diabetic retinopathy using compressed images. Several studies examining the application of JPEG images to medical imaging have shown that relatively low levels of lossy compression (1: 1 to 40: 1) are very accurate in the reproduction of diagnostically useful images. 19 •22 ·25 ·28 In addition, Savcenko and colleagues20 found no substantial difference between images compressed at a ratio of 80: 1 versus 40: 1 using lossy wavelet compression. In contrast, Lee and coworkers25 reported that a high level of JPEG compression (80: 1) of retinal photographs resulted in "pixelated" images of poor quality. This apparent discrepancy can be explained primarily by differences in digitization and viewing techniques. For their study, Lee and coworkers scanned film fundus photographs at a maximum resolution of 2000 dots per inch for subsequent JPEG compression. However, their original TIFF resolution of 2000 dots per inch is misleading because it refers only to pixel density rather than actual information. This discrepancy is indicated by the file size of their TIFF images: a 1.2-MB file implies an image that is only 0.5 megapixels. Compared to our 6-megapixel TIFF images, the digital images used for their study were already poor before undergoing compression. A further source of degradation was introduced by grading colour prints of the scanned images, as the final resolution of the image is limited by the resolution of the slide printer. Our teleophthalmology system avoids both these problems by using a high-resolution image to start with and by viewing the images on a high-resolution monitor equipped with stereoviewing software and hardware. Slight statistical abnormalities were produced in our study as a result of the relatively small sample. K values for venous beading, vitreous or preretinal hemorrhage and very severe NPDR were undefined at both levels of JPEG compression because these abnormalities and this level of diabetic retinopathy were not present in any of the patients. The decision to include only patients presenting to a single teleophthalmology clinic allowed investigation of the appli-

JPEG compression of retinal images-Baker et al

cation of JPEG compression in a realistic environment. Unfortunately, this did not ensure that all types of pathology and levels of retinopathy would be included in our sample, and, as a result, there were fewer patients with severe diabetic retinopathy. It should be noted that no false-positive results were produced by image compression for venous beading, vitreous or preretinal hemorrhage or very severe NPDR, and the rate of exact agreement was high for all three variables (100%). However, future study is needed to prove that JPEG compression will not result in the production of false-negative results, which would prevent clinicians from correctly diagnosing higher levels of diabetic retinopathy. Our results suggest a slight superiority of the images compressed 113x over those compressed 55x. For example, in the identification of intraretinal microvascular abnormalities, good reproducibility (K = 0.46) was obtained with 55x compression, whereas excellent reproducibility (K =0.77) was obtained with 113x compression. Owing to the relatively small number of patients, this difference was not statistically significant. However, this counterintuitive finding suggests the effect of a phenomenon previously described in other studies of image compression. Several investigators have reported that radiologists appear to have a preference for radiographs that have been mildly compressed, and observers often report that the compressed images are of higher quality than the originals. 19 Lee and coworkers25 noted a similar effect of JPEG compression on retinal images: readers reported that fundus images at a compression level of 30: I were selected to be the image of the highest quality 40% of the time. This unintended effect of JPEG compression may result from its underlying mathematical algorithms. Further research will determine whether the finding of increased agreement and reproducibility at levels of compression as high as 113x is a true effect of JPEG compression rather than a statistical artefact. One of the most important aspects of evaluation of diabetic retinopathy following diagnosis is the recommendation for follow-up or referral for treatment to prevent vision loss. 5 Accurate suggestions are especially crucial when using a teleophthalmology system for remote diagnosis rather than simply for screening. Therefore, we needed to ensure that the use of compression would not compromise the follow-up recommendations we made for our patients. We observed a high level of accuracy and reproducibility when com-

paring the original TIFF images to the JPEG compressed photographs. Good to perfect reproducibility was also observed for all follow-up recommendations at both compression levels. The highest levels of reproducibility were observed for the recommendation of immediate referral. It appears that the use of JPEG compression at levels as high as 113:1 should not have any adverse effects on long-term care for patients examined with our system. CONCLUSION

This pilot study was designed to investigate whether JPEG image compression could be used to reduce digital image file sizes while maintaining sufficient quality and detail to accurately diagnose diabetic retinopathy. We have shown that JPEG compression at ratios of 55:1 and 113:1 has a high level of reproducibility for the identification of diabetic retinal abnormalities, levels of diabetic retinopathy and follow-up recommendations when compared to uncompressed TIFF images. We believe that JPEG compression of up to 113: 1 would produce no deleterious influences on teleophthalmologic care of diabetic patients. With the advancements in digital photography, the use of digital imaging in the field of ophthalmology is rapidly expanding. The demands of the large image files produced by new high-resolution digital cameras, however, place an enormous strain on current digital storage and transmission systems. Digital image compression may offer a solution to these problems. REFERENCES

1. Diabetic Retinopathy Study Research Group. Prelimi-

nary report on effects of photocoagulation therapy. Am J Ophthalmoll976;94:505-37. 2. MacMillan HL, MacMillan AB, Offord DR, Dingle JL. Aboriginal health. Can Med Assoc J 1996;155(11): 1569-77.

3. Young TK, Shah C. Extent and magnitude of the problem. In: Young TK, editor. Diabetes in the Canadian native population: bicultural perspectives. Toronto: Canadian Diabetes Association; 1987. p. 11-25.

4. Evers S, McCracken E, Antone I, Deagle G. Prevalence of diabetes in Indians and Caucasians living in southwestern Ontario. Can J Public Health 1987;78(4):240-3.

CAN J OPHTHALMOL-VOL. 39, NO.7, 2004

753

JPEG compression of retinal images-Baker et al

5. Meltzer S, Leiter L, Daneman D, Gerstein HC, LauD, Ludwig S, et al. 1998 clinical practice guidelines for the management of diabetes in Canada. Canadian Diabetes Association. Can Med Assoc J 1998;159 (Suppl8):S1-29. 6. McGregor G. Transmitting digital images of the fundus for the West Virginia Diabetic Eye Care Project. J Ophthalmic Photogr 1994;16:87-98. 7. Young S, George LD, Lusty J, Owens DR. A new screening tool for diabetic retinopathy: the Canon CR5 45NM retinal camera with Frost Medical Software RIS-lite digital imaging system. J Audiov Media Med 1997;20(1): 11--4. 8. Li HK. Telemedicine and ophthalmology. Surv Ophthalmol1999;44(1):61-72. 9. Tennant MTS, Greve MDJ, Rudnisky CJ, Hillson TR, Hinz BJ. Identification of diabetic retinopathy by stereoscopic digital imaging via teleophthalmology: a comparison to slide film. Can J Ophthalmol 2001; 36(4): 187-96. 10. Early Treatment Diabetic Retinopathy Study Research Group. Grading diabetic retinopathy from stereoscopic color fundus photographs - an extension of the modified Airlie House classification. ETDRS report number 10. Ophthalmology 1991;98(5 Suppl):786--806. 11. Wong D. Textbook of ophthalmic photography. New York: Inter-Optics Publications; 1982. p. 76. 12. Early Treatment Diabetic Retinopathy Study Research Group. Fundus photographic risk factors for progression of diabetic retinopathy. ETDRS report number 12. Ophthalmology 1991;98(5 Suppl):823-33. 13. Agresti A. Categorical data analysis. New York: Wiley; 1990. p. 365-70. 14. Rudnisky CJ, Hinz BJ, Tennant MTS, De Leon AR, Greve MDJ. High-resolution stereoscopic digital fundus photography versus contact lens biomicroscopy for the detection of clinically significant macular edema. Ophthalmology 2002;109(2):267-74. 15. Schouten HJ. Estimating kappa from binocular data and comparing marginal probabilities. Stat Med 1993;12(23):2207-17. 16. Fleiss JL. Large sample standard errors of kappa and weighted kappa. Psycho! Bull1969;72:323-7. 17. Oden NL. Estimating kappa from binocular data. Stat Med 1991;10(8):1303-11.

754

CAN J OPHTHALMOL-VOL. 39, NO.7, 2004

18. Rosner B. Fundamentals of biostatistics. 4th ed. Belmont (CA): Wadsworth; 1995. p. 423-8. 19. Smith I, Roszkowski A, Slaughter R, Sterling D. Acceptable levels of digital image compression in chest radiology. Australas Radiol2000;44(1):32-5. 20. Savcenko V, Erickson BJ, Palisson PM, Persons KR, Manduca A, Hartman T, et al. Detection of subtle abnormalities on chest radiographs after irreversible compression. Radiology 1998;206(3):609-16. 21. Strintzis MG. A review of compression methods for medical images in PACS. Int J Med Inform 1998; 52:159-65. 22. Slone RM, Foos DH, Whiting BR, Muka E, Rubin DA, Pilgram TK, et al. Assessment of visually lossless irreversible image compression: comparison of three methods by using an image-comparison workstation. Radiology 2000;215(2):543-53. 23. MacMahon H, Doi K, Sanada S, Montner SM, Giger ML, Metz CE, et al. Data compression: effect on diagnostic accuracy in digital chest radiography. Radiology 1991;178(1):175-9. 24. Cosman PC, Davidson HC, Bergin CJ, Tseng CW, Moses LE, Riskin EA, et al. Thoracic CT images: effect of lossy image compression on diagnostic accuracy. Radiology 1994;190(2):517-24. 25. Lee MS, Shin DS, Berger JW. Grading, image analysis, and stereopsis of digitally compressed fundus images. Retina 2000;20(3):275-81. 26. Nelson M. The data compression book. New York: M & T Publishing; 1992. p. 347-73. 27. Anagnoste SR, Orlock D, Spaide R, Guyer DR, Slakter GJ, Sorenson J, et al. Digital compression of ophthalmic images: maximum JPEG compression undetectable by retinal specialists [abstract]. Invest Ophthalmol Vis Sci 1998;39:S922. 28. Baker WA, Hearne SE, Spero LA, Morris KG, Harrington RA, Sketch MH, et al. Lossy (15: 1) JPEG compression of digital coronary angiograms does not limit detection of subtle morphological features. Circulation 1997;96(4):1157-64. Key words: diabetic retinopathy, diabetes mellitus, retina, telemedicine, fundus photography, medical imaging, teleophthalmology