Impact of JPEG lossy image compression on ... - BIR Publications

3 downloads 0 Views 251KB Size Report
Objectives: The aim of the study was to evaluate the impact of JPEG lossy image compression on the estimation of alveolar bone gain by quantitative digital ...
ã

Dentomaxillofacial Radiology (2002) 31, 106 ± 112 2002 Nature Publishing Group. All rights reserved 0250 ± 832X/02 $25.00 www.nature.com/dmfr

RESEARCH

Impact of JPEG lossy image compression on quantitative digital subtraction radiography A Fidler1, B Likar2, F PernusÆ 2 and U SkalericÆ*,3 1

University of Ljubljana, Faculty of Medicine, Department of Stomatology, Ljubljana, Slovenia; 2University of Ljubljana, Faculty of Electrical Engineering, Department of Systems Control and Cybernetics, Ljubljana, Slovenia; 3University of Ljubljana, Faculty of Medicine, Department of Oral Medicine & Periodontology, Ljubljana, Slovenia

Objectives: The aim of the study was to evaluate the impact of JPEG lossy image compression on the estimation of alveolar bone gain by quantitative digital subtraction radiography (DSR). Methods: Nine dry domestic pig mandible posterior segments were radiographed three times (`Baseline', `No change', and `Gain') with standardized projection geometry. Bone gain was simulated by adding arti®cial bone chips (1, 4, and 15 mg). Images were either compressed before or after registration. No change areas in compressed and subtracted `No change ± Baseline' images and bone gain volumes in compressed and subtracted `Gain ± Baseline' images were calculated and compared to the corresponding measurements performed on original subtracted images. Results: Measurements of no change areas (`No change ± Baseline') were only slightly a€ected by compressions down to JPEG 50 (J50) applied either before or after registration. Simulated gain of alveolar bone (`Gain ± Baseline') was underestimated when compression before registration was performed. The underestimation was bigger when small bone chips of 1 mg were measured and when higher compression rates were used. Bone chips of 4 and 15 mg were only slightly underestimated when using J90, J70, and J50 compressions before registration. Conclusions: Lossy JPEG compression does not a€ect the measurements of no change areas by DSR. Images undergoing subtraction should be registered before compression and if so, J90 compression with a compression ratio of 1 : 7 can be used to detect and measure 4 mg and larger bone gain. Dentomaxillofacial Radiology (2002) 31, 106 ± 112. DOI: 10.1038/sj/dmfr/4600670 Keywords: diagnostic imaging, subtraction technique; radiography, dental; image compression; alveolar process Introduction Digital imaging is now an established technique for dentistry with typical intraoral radiographic images requiring 100 to 300 kB of storage space.1 ± 2 Transmission of images between clinicians or to insurance payers is likely to increase in the future and internet bandwidth may become a limiting factor. Reduction of image ®le size by compression may be necessary to ensure rapid transmission of data. There are two basic types of image data compression: lossless and lossy. After lossless compression, typically a 1 : 2 reduction in ®le size, the *Correspondence to: Prof Dr U SkalericÆ, Department of Oral Medicine & Periodontology, Hrvatski trg 6, SI-1000 Ljubljana, Slovenia; E-mail: [email protected] Received 2 July 2001; revised 16 October 2001; accepted 25 October 2001

original image can be recovered. In lossy compression the ®le size reduction can be much greater (1 : 10 or more) but at the expense of some loss of the original data which may reduce diagnostic accuracy. Several studies have shown that lossy image compression does not signi®cantly a€ect the diagnostic accuracy in medical3 ± 6 and dental7 ± 11 radiology, such as a 95% intraoral image compression for caries diagnosis.10 While there have been a number of studies that evaluated the quality of images recovered from the compressed format, the e€ect of lossy compression on subsequent image analysis, such as feature extraction or volume determination, has received little attention. Lossy image compression can only recover an approximation of the original image and may, there-

JPEG compression and DSR A Fidler et al

fore, adversely a€ect any subsequent automatic image analysis task. For example, digital subtraction radiography (DSR) is a valuable tool for detecting small changes in mineralisation over time.12 ± 15 However, a prerequisite for DSR is that images match closely in projection geometry and contrast, which requires a special acquisition protocol. Small residual mismatches in projection geometry and contrast can and should be corrected for retrospectively by image registration.16 Changes in image data induced by lossy image compression may, besides directly a€ecting the images to be subtracted, hamper the registration and thereby indirectly generate additional false positive or false negative changes in the subtracted image. The aim of this study was to assess the impact of JPEG lossy image compression on quantitative digital subtraction radiography. Materials and methods Ten dry posterior jaw segments were obtained from ®ve domestic pig mandibles, but due to damage only nine segments were used. The teeth were ®xed into their sockets by cyanoacrylic glue (Pattex, Henkel, DuÈsseldorf, Germany). Projection geometry was standardized by an optical bench consisting of X-ray machine cone seat, mandible section holder, and receptor plate holder (Figure 1). The mandible section holder had a rectangular impression with convergent walls. Each section of mandible was ®xed to a block of dental stone on its lower end, which ®tted tightly into the rectangular impression and thus provided repeatable projection geometry. The X-ray machine (Prostyle Intra, Planmeca, Helsinki, Finland) was operated at 70 kV, 8 mA, and 0.12 s exposure time. All nine sections were radiographed three times on a Digora (Soredex Orion Corporation, Helsinki, Finland) imaging plate (30640 mm). The ®rst two series `Baseline' and `No change' were radiographed under the same conditions,

Figure 1 Optical bench for standardization of projection geometry with (1) receptor plate holder, (2) mandible section holder, and (3) Xray cone seat

while before the third series `Gain' was radiographed, a piece of ®lm with nine arti®cial hydroxyapatite bone chips (Ceros 80, Mathys Co., Bettlach, Switzerland) was inserted between a mandible and receptor plate holder. Three small, three medium, and three large bone chips, weighting 1, 4, and 15 mg, respectively, were attached to the ®lm. The shape of bone chips was approximately oval with irregular edges (Figure 2). Storage phosphor plates were scanned and images (6286466 pixels, 256 gray levels, 64664 mm pixel size) were saved on hard disk. All 27 images, nine `Baseline', nine `No change', and nine `Gain', were duplicated into sets named A and B. Images in set A were ®rst compressed/decompressed, then registered, and ®nally subtracted, while images in set B were ®rst registered, then compressed/decompressed, and ®nally subtracted. Images were compressed/decompressed by Microsoft Photo Editor 3.0 (Microsoft Corp., Redmond, WA, USA), using four di€erent compression levels, i.e., J90, J70, J50 and J30, which equidistantly sampled the compression scale. `Baseline' images were used to calculate average ®le sizes of original images, compressed images, and compression ratios at different compression levels. Registration between corresponding `No change ± Baseline' and `Gain ± Baseline' images was carried out by a non-parametric contrast correction algorithm17 and a projective geometric transformation obtained by maximization of mutual information.18,19 In contrast correction, the intensity values in `No change' and `Gain' images were changed so that their cumulative density functions matched the cumulative density functions of the

107

Figure 2 Uncompressed `Baseline' image (top left), `Gain' image (top right), corresponding subtracted image (bottom left), and ®lm with attached bone chips (bottom right). Bright rectangle in each image illustrates a region of interest containing a 4 mg bone chip Dentomaxillofacial Radiology

JPEG compression and DSR A Fidler et al

108

corresponding `Baseline' images.17 In geometric registration, projective transformations were applied to the `No change' and `Gain' images so that the similarities between the corresponding `Baseline' images were maximized. Mutual information was used as a similarity criterion because it had been proven robust and ecient in various registration tasks.18,19 Registered `No change ± Baseline' and `Gain ± Baseline' images in sets A and B were ®nally subtracted. The median gray value of 128, added to each subtracted image, represented no di€erence. For each bone chip a square region of interest (ROI) was manually selected (Figure 2). One large bone chip was excluded from evaluation because it was not always lying completely inside the image domain. There were thus three small, three medium, and two large ROIs, with a side of 40, 55 and 80 pixels, respectively. By analysing the original `No change ± Baseline' subtracted images, it was found that 99% of pixels in ROIs had a gray value between 122 and 134, i.e. within +3 standard deviations. Therefore, all pixels in the subtracted images with gray value less than 122 or more than 134 were considered as `change'. To evaluate the in¯uence of compression, applied either before or after registration, on `No change ± Baseline' subtractions, the percentages of pixels showing `no change' in ROIs were determined for each compression level and compared to values obtained from corresponding uncompressed subtractions. The in¯uence of compression on bone gain measurements obtained by `Gain ± Baseline' subtractions was expressed by the volume of bone gain, which was de®ned as the weighted sum of all pixels in ROIs that showed change. A weight represented the absolute gray level di€erence between the pixel gray value and the no di€erence value of 128. The volumes that had been de®ned for compressed `Gain ± Baseline' images in sets A and B and uncompressed `Gain ± Baseline' images were used to calculate relative volumes of bone gain. For statistical analysis Friedman's two-way analysis of variance by ranks and Wilcoxon signed rank test were used. A P-value of less than 0.05 was considered statistically signi®cant. Statistical analysis was performed by SPSS 9.0.0 package for Windows (SPSS Inc., Chicago, Illinois, USA). The results were illustrated by box-whiskers diagrams that showed minima, 1st quartiles, medians, 3rd quartiles, and maxima of percentages of no change or relative volumes of bone gain. Results Compression levels and compression ratios The average ®le sizes and corresponding compression ratios for di€erent compression levels are presented in Table 1. The highest relative reduction in ®le size was 247 kb, obtained by J90. Further compressions resulted in less signi®cant relative reductions of ®le sizes. To

Dentomaxillofacial Radiology

Table 1 Average file sizes for different compression levels Compression level Original J90 J70 J50 J30

File size (% of original)

File size (kB)

Compression ratio

100.0 14.1 6.4 4.5 3.3

288 41 19 13 9

1:1 1:7 1 : 16 1 : 22 1 : 31

qualitatively assess the e€ect of compression on the original and subtracted images from sets A and B, magni®ed region of interest depicted in Figure 2 is shown in Figure 3. Compression artifacts, which are larger when higher compression rates are applied, can be seen in `Baseline' and `Gain' images and even better in corresponding subtractions. Quantitative results on the impact of image compression are given in the following. Impact of compression on `No change ± Baseline' DSR Figure 4 illustrates the percentages of pixels that showed `no change' in ROIs of subtracted `No change ± Baseline' images. No statistically signi®cant differences in percentages of true `no change' pixels in ROIs were found between original uncompressed images and images in set A, except with compression J30, which yielded smaller percentages of true `no change' pixels (P50.001). In set B, the di€erences of percentages of true `no change' pixels in ROIs found after J90, J70 and J30 compressions were signi®cant (P50.001), although the median di€erences were very subtle. After applying the J50 compression, the di€erences found were not statistically signi®cant. Impact of compression on `Gain ± Baseline' DSR Figures 5, 6 and 7 show the impact of compression on relative volumes of bone gain for small, medium, and large bone chips, respectively. Small bone chips (1 mg): For images compressed before registration (set A), the volumes were underestimated (P50.001) after J90, J70 and J50 compressions (Figure 5). The corresponding medians of relative volumes were 52.6, 48.9 and 66.7%. After J30 compression, the median relative volume was 95.0% and was not statistically signi®cant. The J90 compression yielded a 95.9% median relative volume (P=0.020), when images were compressed after registration (set B). For other compression levels the bone gain volumes were not statistically signi®cantly di€erent from volumes obtained from the corresponding uncompressed subtracted images. The interquartile range (1st ± 3rd) was increasing with compression level. Medium bone chips (4 mg): In set A, volumes were signi®cantly (P50.001) underestimated (Figure 6). The median relative volumes were 78.7, 76.2 and 79.5 for

JPEG compression and DSR A Fidler et al

109

Figure 3 Magni®ed region of interest for 4 mg bone chip, depicted in Figure 2, illustrating the e€ect of di€erent compression levels on `Baseline' and `Gain' images and corresponding subtractions

J90, J70 and J50 compressions, respectively. After J30 compression, the relative bone gain volume found was 91.6%, which was not statistically signi®cant. In set B, the volumes were only slightly underestimated after J90, J70 and J50 compressions and slightly overestimated after J30. The corresponding median relative volumes were 98.4, 96.1, 96.7 and 103.3%. Interquartile range was again increasing with compression level but not so dramatically as in set A. Large bone chips (15 mg): In set A, volume underestimation (P50.001) was found after J90, J70 and J50 compression (Figure 7). The medians of relative volumes were 83.4, 82.4 and 84.8%, respectively. After J30 compression, the obtained relative volume

was 93.8% (P=0.012). In set B, negligible underestimation of volumes after J90, J70 and J50 compression and overestimation after J30 were found. Corresponding relative median volumes were 99.8, 99.9, 98.4 and 100.5%. Interquartile range was slightly increasing with compression level, from 1.2% for J90 to 5.6% for J30. Discussion Lossy JPEG image compression is utilizing a known perceptual characteristic of the human eye to achieve higher compression and preserve the most relevant information content. Even if `perceptually lossless', i.e. Dentomaxillofacial Radiology

JPEG compression and DSR A Fidler et al

110

Figure 4 Percentage of pixels in ROIs showing `no change' in subtracted `No-change ± Baseline' images given for the original uncompressed images (O) and for di€erent compression levels and both experimental sets (A and B). #- statistically signi®cant (P50.05)

does not hamper the visual interpretation of digital radiographs, JPEG compression may a€ect computer analysis of digital images. The possible adverse e€ects of compression in a speci®c application domain can only be assessed by experimentation. In this study, we investigated the impact of JPEG compression on assessing alveolar bone gain by digital subtraction radiography. No large di€erences were found between the subtracted original `No change ± Baseline' images and corresponding `No change ± Baseline' images compressed by J90, J70 or J50 either before or after registration. These ®ndings thus indicate that lossy JPEG compression did not induce false positive bone gain as long as compression level was below J30. Moreover, the results for J90 and J70 compressed images in set B show, that the percentage of no change area was even slightly greater than in the corresponding original subtractions. These results are in accordance with results of Janhom and coworkers10, who also found that JPEG reduces noise and thereby reduces false positive bone gain. The second observation is that lossy JPEG compression may adversely a€ect the quanti®cation of the true positive bone gain, especially if images are compressed before registration (set A). The use of JPEG compression is re¯ected in Dentomaxillofacial Radiology

Figure 5 Relative volumes of bone gain, given for sets A and B and for di€erent compression levels, obtained from ROIs containing small bone chips (1 mg). #- statistically signi®cant (P50.05)

underestimated bone gain. This observation leads to an important conclusion that lossy image compression hampers the registration process, which is composed of contrast and geometric matching of images. The ®nding suggests that image registration should precede image compression. As one can expect, the higher the compression ratio and the smaller the bone gain is, the less accurate and reliable the measurements will be. In this study, it turned out that J90 compression, having a compression ratio of 1 : 7, yields acceptable accuracy and reliability for a large range of bone gain (4 to 15 mg), indicating that lossy J90 compression can be used for DSR if only registration precedes compression (set B). Reliable quanti®cation of 1 mg bone gain is still a challenge in DSR also without image compression.20 The obtained acceptable compression ratio of 1 : 7 for DSR is nearly twice smaller than the acceptable compression ratio for conventional visual diagnosis.8 The above observations are based on results from a fairly well-controlled in vitro experiment, using standardization of projection geometry and arti®cial bone chips. Although such an experiment could not completely mimic the actual variability of projection geometry and bone gain, it provided a `gold standard' image database that was useful for quantitative evaluation of image compression in DSR, which was

JPEG compression and DSR A Fidler et al

111

Figure 6 Relative volumes of bone gain, given for sets A and B and for di€erent compression levels, obtained from ROIs containing medium bone chips (4 mg). #- statistically signi®cant (P50.05)

Figure 7 Relative volumes of bone gain, given for sets A and B and for di€erent compression levels, obtained from ROIs containing large bone chips (15 mg). #- statistically signi®cant (P50.05)

our primary goal. In an alternative study, one could conduct a clinical evaluation with a more representative image database. However, because of the missing `gold standard', objectivity, and evaluation criteria, the relevance of the results to DSR would be questionable. Another problem with assessing the impact of image compression is the inconsistency of the compression scale between di€erent compression methods and/or compression software. Already within the JPEG compression there is an inconsistency between the compression scales, which makes the comparison between studies published so far dicult and calls for a standardization of the compression levels. Yet, when comparing di€erent compression methods, compression level cannot be used, as there is no direct relation between the compression levels of di€erent methods because actual compression is image- and method-dependent. In this case, `compression ratio' or `number of bits per pixel' are the parameters of choice for further studies that would enable better comparison. Besides, validation of compression software implementation is another important issue that was already encountered in studying the e€ect of di€erent image ®le formats and image analysis software programs on dental radiometric digital evaluations.21

In conclusion, the usefulness of image compression depends critically on the quality of the processed images, where image quality is an attribute with many possible de®nitions and interpretations, depending on the use to which the images will be put. For this reason, the measurement of image quality is a dicult task that was quantitatively considered by only a few researchers and no single approach to qualitatively measure it has gained universal acceptance.11 We conducted a speci®c evaluation of lossy JPEG compression for quantifying bone gain by digital subtraction in dental radiography. The obtained results indicated that, ®rst, lossy JPEG compression does not a€ect the detection of no bone change areas. Second, lossy image compression hampers the registration process, which suggests that image registration should generally precede image compression. And ®nally, it turned out that J90 compression with compression ratio of 1 : 7, yields acceptable accuracy and reliability for detecting and measuring 4 mg or larger bone gain. Acknowledgements The authors would like to thank Braintec d.o.o., Ljubljana, Slovenia, for material support. Dentomaxillofacial Radiology

JPEG compression and DSR A Fidler et al

112

References 1. Versteeg CH, Sanderink GCH, van der Stelt PF. Ecacy of digital intra-oral radiography in clinical dentistry. J Dentistry 1997; 25: 215 ± 224. 2. Wong S, Zaremba L, Gooden D, Huang HK. Radiographic image compression ± A review. Proc. IEEE 1995; 83: 194 ± 219. 3. MacMahon H, Doi K, Sanada S, Montner SM, Giger ML, Metz CE, et al. Data compression: E€ect on diagnostic accuracy in digital chest radiography. Radiology 1991; 178: 175 ± 179. 4. Good WF, Maitz GS, Gur D. JPEG compatible data compression of mammograms. J Digit Imaging 1994; 7: 123 ± 132. 5. Sneiderman C, Schosser R, Pearson TG. A comparison of JPEG and FIF compression of color medical images for dermatology. Comput Med Imaging Graph 1994; 18: 339 ± 342. 6. Baker WA, Hearne SE, Spero LA, Morris KG, Harrington RA, Sketch Jr MH, et al. Lossy (15 : 1) JPEG compression of digital coronary angiograms does not limit detection of subtle morphological features. Circulation 1997; 96: 1157 ± 1164. 7. Wenzel A, Borg E, Hitze H, GroÈndahl HG. Accuracy of caries diagnosis in digital images from charge-coupled device and storage phosphor system: an in vitro study. Dentomaxillofac Radiol 1995; 24: 250 ± 254. 8. Wenzel A, Gotfredsen E, Borg E, GroÈndahl HG. Impact of lossy image compression on accuracy of caries detection in digital images taken with a storage phosphor system. Oral Surg Oral Med Oral Pathol Oral Radiol Endod 1996; 81: 351 ± 355. 9. Yuasa H, Ariji Y, Ohki M, Naitoh M, Shiojima M, Ushida M, et al. Joint Photographic Experts Group compression of intraoral radiographs for images transmission on the World Wide Web. Oral Surg Oral Med Oral Pathol Oral Radiol Endod 1999; 88: 93 ± 99. 10. Janhom A, van der Stelt PF, van Ginkel FC. Interaction between noise and ®le compression and its e€ect on the recognition of caries in digital imaging. Dentomaxillofac Radiol 2000; 29: 20 ± 27. 11. Cosman P, Gray R, Olshen R. Quality evaluation for compressed medical images: Fundamentals. In: Bankman (ed). Handbook of Medical Imaging. San Diego: Academic Press, 2000, pp 803 ± 819.

Dentomaxillofacial Radiology

12. GroÈndahl HG, GroÈndahl K. Subtraction radiography for the diagnosis of periodontal bone lesions. Oral Surg Oral Med Oral Pathol 1983; 55: 208 ± 213. 13. Vos MH, Janssen PT, Van Aken J, Heethaar RM. Quantitative measurement of periodontal bone changes by digital subtraction. J Periodontal Res 1986; 21: 583 ± 591. 14. BraÈgger U, Pasquali L, Rylander H, Carnes D, Kornman KS. Computer-assisted densitometric image analysis in periodontal radiography. A methodological study. J Clin Periodontol 1988; 15: 27 ± 37. 15. Je€coat MK, Reddy MS. Digital subtraction for longitudinal assessment of peri-implant bone change: method and validation. Adv Dent Res 1993; 7: 196 ± 201. 16. Lehmann TM, GroÈndahl HG, Benn DK. Computer-based registration for digital subtraction in dental radiology. Dentomaxillofac Radiol 2000; 29: 323 ± 346. 17. Ruttimann UE, Webber RL, Schmidt E. A robust digital method for ®lm contrast correction in subtraction radiography. J Periodontal Res 1986; 21: 486 ± 495. 18. Likar B, PernusÆ F. A hierarchical approach to elastic registration based on mutual information. Image and Vision Computing 2001; 9: 33 ± 44. 19. Likar B, PernusÆ F. Registration of serial transverse sections of muscle ®bers. Cytometry 1999; 37: 93 ± 106. 20. Hausmann E, Allen K, Loza J, Buchanan W, Cavanaugh PF. Validation of quantitative digital subtraction radiography using the electronically guided alignment device/impression technique. J Periodontol 1996; 67: 895 ± 899. 21. GuÈrdal P, Hildebolt CF, Akdeniz BG. The e€ects of di€erent image ®le formats and image-analysis software programs on dental radiometric digital evaluations. Dentomaxillofac Radiol 2001; 30: 50 ± 55.