Dentomaxillofacial Radiology (2002) 31, 257 ± 263 ã 2002 Nature Publishing Group. All rights reserved 0250 ± 832X/02 $25.00 www.nature.com/dmfr
RESEARCH
A comparison of two compression algorithms and the detection of caries A Janhom*,1,2, PF van der Stelt1 and GCH Sanderink1 1
Department of Oral Radiology, Academic Centre for Dentistry Amsterdam (ACTA), Amsterdam, The Netherlands; 2Department of Oral Radiology, Faculty of Dentistry, Chiang Mai University, Chiang Mai, Thailand
Objectives: To assess the eect of two compression algorithms (JPEG and wavelet) on the detection of approximal caries. Methods: Fifteen bitewing radiographs were generated using 100 posterior teeth mounted in blocks. The images were produced on conventional ®lms (Ektaspeed Plus) and scanned at 300 d.p.i. Digital images were then compressed 9 : 1 with JPEG and wavelet methods. Nine observers detected the presence and depth of approximal caries recorded on a 5-point con®dence scale and a 4-point depth scale from images viewed in random order. Histological examination provided the true depth of the lesions. Data were analysed by means of ANOVA. The null hypothesis was that there is no signi®cant dierence between the two compression algorithms and the original uncompressed images. Results: JPEG performed signi®cantly worse than the original and the wavelet algorithm (P50.001) for the detection of dentinal lesions. However, no signi®cant dierences were found for the detection of sound surfaces, enamel lesions, and lesions up to the DEJ between JPEGcompressed images and each of the other two modalities. There was also no signi®cant dierence between the wavelet-compressed images and the original for all lesion depths. Conclusions: At a compression ratio of 9 : 1, there were no signi®cant dierences among the original images, JPEG and wavelet compressed images for the detection of enamel caries. JPEG-compressed images performed inferiorly to the original and wavelet-compressed images for the detection of dentinal lesions. Wavelet compression is a better choice than JPEG at the compression ratio investigated in this study. Dentomaxillofacial Radiology (2002) 31, 257 ± 263. doi:10.1038/sj.dmfr.4600704 Keywords: image processing, computer-assisted; digital radiography, image compression; observer performance; dental caries Introduction Digital radiography has many advantages, including the transmission of images for consultation or for electronic claims submission and the use of image compression algorithms can reduce ®le size and transmission times.1 Compression algorithms can be categorized into `Lossless' or `Lossy' where the former method preserves the information of the original image pixel-by-pixel. Usually lossless images cannot achieve more than a 3 : 1 compression ratio. However, lossy
*Correspondence to: Apirum Janhom, ACTA, Department of Oral Radiology, Louwesweg 1, 1066 EA Amsterdam, The Netherlands; E-mail:
[email protected] Received 9 April 2001; revised 19 June 2001; accepted 16 July 2001
methods can achieve higher compression ratios but there is some loss of information although this may not be apparent to the human eye. Examples of lossy algorithms are JPEG (Joint Photographic Expert Group), wavelet and fractal techniques. The JPEG lossy compression algorithm has been adopted as a standard for compression by the Digital Imaging and Communications in Medicine (DICOM) committee. While many studies have been conducted on the use of compression algorithms in medical research2 ± 10, little is known about the eect of JPEG compression on digital radiographs used in dentistry. Van der Stelt et al.11 reported that a JPEG compression ratio of 25 : 1 did not reduce the diagnostic utility of endodontic ®le tip images. Sanderink et al.12 reported
Compression algorithms A Janhom et al
258
that panoramic images can be reduced 28 : 1 in size by JPEG compression before the detection of radiolucencies in bone is signi®cantly aected. For caries diagnosis, compression ratios of 12 : 1 and 14 : 1 were reported without deterioration of diagnostic accuracy.13,14 However, recent investigations have demonstrated that wavelet-based methods oer signi®cantly better quality at the same ratio than JPEG methods for medical image compression.5 ± 10 Only one study reports on wavelet compression in dentistry where temporomandibular joint images were studied.15 Wavelet compression has not been investigated for other imaging applications in dentistry. The aim of this study is to assess the performance of two lossy compression algorithms (JPEG and wavelet) at the same level of compression and to compare them with the original images for the detection of caries. Materials and methods Image acquisition Bitewing radiographs were generated using 100 extracted posterior teeth mounted in blocks: 52 premolars and 48 molars. Premolar teeth were mounted together in sets of four and molars in groups of three. In total, 13 premolar blocks and 16 molar blocks were constructed with full details available in our previous study.19 A total of 15 radiographs simulating the bitewing projection technique were obtained on conventional No. 2 size Ektaspeed Plus ®lms (Eastman Kodak, Rochester, NY, USA). A Heliodent MD dental X-ray machine (Siemens, Bensheim, Germany) was used at 60 kVp, 7 mA, a FFD of 35 cm and rectangular collimation. A 12-mm thick soft tissue-equivalent material (Mix-D)16,17 was placed between the X-ray source and the tooth blocks to simulate the scattering eect of soft tissue. The exposure time was 0.32 s for the premolar blocks and 0.40 s for the molars. Using these settings, the density of dentine close to the proximal enamel was on average 0.80 when measured with a densitometer (MacBeth TD 502, Newburg, NY, USA). This is considered a clinically acceptable density.18 The radiographs were developed in an automatic processor (DuÈrr-Dental, Bietigheim-Bissingen, Germany) at 278C, with a 5.4 min processing cycle. Films were checked after having been processed with regard to the density values (maximum density, average density of dentine) to assure constant exposure and processing conditions. Scanning and compression methods The radiographs were masked and secured in ®lm mounts (Dentsply Rinn, Elgin, Il, USA). All radiographs were then scanned on a ¯atbed scanner (Agfa DuoScan T1200; Agfa, Mortsel, Belgium), using the manufacturer's software (Agfa Fotolook 3.0; Agfa, Mortsel, Belgium) at eight bits and resolution set at
Dentomaxillofacial Radiology
300 d.p.i. This resolution was chosen because in a previous study it was shown to be suitable for caries detection19 and is similar to values used for intra-oral radiography storage phosphor plate scanners. The maximum density range of the scanner was D=3.4 and the gamma was set at 2.2, which proved to reproduce all gray levels of the original radiographs. The ®lms were placed in the upper center of the glass surface of the scanner and the image ®les were saved in the BMP format. In total, 200 proximal surfaces were scanned and the ®les compressed with the JPEG program set at level 27 on a scale of 1 ± 255 resulting in a compression ratio of 9 : 1 (Leadtools, Technologies Inc, Charlotte, NC, USA). This value was selected because it was previously reported to be acceptable for caries detection.14 Images were also compressed at the same compression ratio of 9 : 1 with a wavelet algorithm (ViewMed, Pegasus Imaging Corp, Tempa, FA, USA). Subtraction images were produced from the original and the compressed-images, to show the dierence between both algorithms and the characteristics of the image data loss (Figure 1). Observation sessions Images were randomly presented to nine observers who viewed one image at a time in random order. The image modalities (JPEG, wavelet and original) were randomly assigned for each observer. Each image modality was viewed in a dimly lit room as a separate session in order to minimize fatigue and learning eects. The images were viewed 50 cm from a 15-in ¯at panel TFT monitor (LR15; NEC, Korea) with 10246768 pixel resolution and `true color' setting. All nine observers were either sta members (full-time or part-time) or graduate students at the departments of Oral and Maxillofacial Radiology or Cariology Endodontology Pedodontology at the Academic Centre for Dentistry Amsterdam. The presence of caries was scored on a 5-point scale: 1=de®nitely no proximal radiolucency, 2=probably no radiolucency, 3=uncertain, 4=probably radiolucency and 5=de®nitely radiolucency. For each score of four or ®ve, they were asked to specify the depth of the lesion as follows: 1=radiolucency restricted to the outer half of enamel, 2=radiolucency extending to the Dentino-Enamel Junction (DEJ) and 3=radiolucency extending deep into dentine. Histological examination of the lesion depth After the observation sessions, all teeth were sectioned mesio-distally close to the lesion using a low speed diamond bur, and ground so that the deepest part of the lesion could be histologically examined with a stereomicroscope with 106 magni®cation (Wild 355110, Heerbrugg, Switzerland). Consensus of the histological results was obtained from two investigators. An enamel lesion was de®ned when it was limited to the outer half of the enamel, a DEJ lesion when it
Compression algorithms A Janhom et al
reached as far as the DEJ and not more than 0.5 mm into the dentine, and as a dentinal lesion when it extended deeper into dentine. Some surfaces (n=30) were damaged or found to have root caries or occlusal caries and were excluded from the study. The ®nal material consisted of 170 approximal surfaces. The distribution of lesion depths based on the histological examination of the sectioned teeth are presented in Table 1.
259
Data analysis The assessments of the nine observers for each surface during the three viewing sessions (JPEG, wavelet, and original) were compared with the results from the histological examination of the sectioned teeth. An Analysis of Variance was performed to compare the two algorithms. The signed and absolute observer errors were calculated. The signed observer error was obtained by subtracting the reference depth from the depth reported by the observer and is a measure to determine the amount of over- or under-estimation of the depth of lesions. However, as a measure of observers' competence it is less adequate because the eects of over- and under-estimation of lesions can cancel each other out. Therefore, the absolute observer error was calculated as well, which is the absolute value of the signed observer error; it indicates the magnitude of the error irrespective of over- or under-estimation. The observer con®dence score was also calculated, which indicates the amount of con®dence expressed (0=uncertain, 1=probably right, 2=de®nitely right). A minus sign was added to the con®dence score when the observer was wrong. A multivariate analysis of variance (MANOVA) was carried out with the signed observer error, absolute observer error and observer con®dence level as dependent variables. The image modality (JPEG, wavelet, and original), and the identity of the observer (1 to 9) were entered as within subjects factor. The histological lesion depth was treated as between subjects factor. Results were considered signi®cant for P50.05. The null hypothesis was that there is no signi®cant dierence among the two compression methods and the original images in caries diagnosis. Data were analysed with the aid of the Statistical Package for the Social Sciences (SPSS version 9.0 for Windows). Table 1 The distribution of lesion depth from histological examination Figure 1 Examples of the original and compressed images by JPEG and wavelet at a compression ratio of 9 : 1, and the respective subtracted images. The contrast enhanced subtraction images show image information loss from the compression procedure with the JPEG loss correlated with image details, while wavelet compression results in more randomly distributed loss of information
Lesion depth by histology
Total
None Enamel DEJ Dentine Total
78 40 31 21 170 Dentomaxillofacial Radiology
Compression algorithms A Janhom et al
260
Results The mean and standard deviation (s.d.) of pixel gray values of the subtraction images of original and JPEGcompressed images were 126.90 s.d. 3.30; for the wavelet subtraction images these values were 127.00 s.d. 3.00 respectively. Figure 1 shows an example of contrast enhanced subtraction images (to emphasise the dierences) of an original radiograph and respectively the JPEG- and the wavelet-compressed versions of this image. The subtraction images represent the information lost during the compression process and shows that the information discarded is more randomly distributed over the image in the wavelet compression than it is in the JPEG-compression. This could indicate that wavelet compression is less detrimental to image details than JPEG compression. However, the test showed that there was no statistically signi®cant dierence of the s.d. of subtraction images between the two compression algorithms. (P50.001). Table 2 shows the outcome of multivariate tests for all three dependent variables (signed observer error, absolute observer error and observer con®dence) together. Among others, there was a signi®cant dierence between the two compression algorithms and the original (P50.001). Further tests revealed that the dierences were in the detection of dentinal lesions
for both signed observer error and absolute observer error (Table 3). Pair-wise comparisons showed that JPEG performed inferiorly to wavelet and the original images in the detection of dentinal lesions (Table 4). No signi®cant dierence was found between the performance of wavelet-compressed images and the original images at any lesion depths. All three modalities performed equally well for the detection of sound surfaces and lesions up to the DEJ. Figure 2 shows the absolute observer error scores for the three modalities at dierent lesion depths. There was no statistically signi®cant dierence in observer con®dence among the three modalities at any lesion depth. The interaction eects of lesion depth and observer were also signi®cantly dierent. This indicates that dierent observers have a dierent response to the compressed images at dierent lesion depths. The reliability calculatzion for the nine observers showed a high inter-observer agreement for each of the
Table 4 Pair-wise comparisons of the three modalities for dentinal lesions. JPEG-compressed images performed significantly inferiorly to wavelet-compressed images and the original
Measure Table 2 The outcome of multivariate tests for all three variables (signed observer error, absolute observer error and observer confidence) together. Type of teeth = premolar, molar. Lesion depth = none, outer enamel, DEJ and dentine. Modalities = JPEG, wavelet and original. Observer = observer 1 ± 9 Sig.a
Parameters Type of teeth Lesion depth Modalities Modality * Type of teeth Modalities * Lesion depth Modalities * Type of teeth * Lesion depth Observer Observer * Lesion depth Modalities * Observer Modalities * Observer * Lesion depth a
Lesion
JPEG vs Wavelet Absolute JPEG vs observer Dentine Original error Wavelet vs Original
Average of absolute error
Difference
1.131 vs 0.213 0.918 1.131 vs 0.191 0.939 0.918 vs ±2.12E-02 0.939
Std. Error of the difference Sig.a 0.038
50.001
0.044
50.001
0.039
0.584
a
Bold print indicates signi®cance at the 0.05 level
0.066 50.001 50.001 0.226 50.001 0.304 50.001 50.001 50.001 50.001
Bold print indicates signi®cance at the 0.05 level
Table 3 Simple resolution within lesion effects test. The significant differences were found in signed observer error and absolute observer error in the detection of dentinal lesions Lesion
Sig.a
Signed observer error
None Enamel DEJ Dentine
0.977 0.123 0.257 50.001
Absolute observer error
None Enamel DEJ Dentine
0.975 0.970 0.246 50.001
Measure
a
Bold print indicates signi®cance at the 0.05 level
Dentomaxillofacial Radiology
Figure 2 Average absolute observer error for all observers at dierent lesion depths for the three modalities (JPEG, wavelet and original). Only the dierence between JPEG and both wavelet and original image for dentinal lesions was signi®cant
Compression algorithms A Janhom et al
modalities (Table 5). Among the nine participants, ®ve of them performed equally well in diagnosing approximal caries; on average their absolute observer error scores were at the same level (Figure 3). Discussion This in vitro study compared two compression methods (JPEG and wavelet) at the compression ratio 9 : 1 with respect to the detection of approximal caries. Theoretically, the performance of the two compression algorithms at this low compression rate should be similar. Slone et al.4 found in their study on chest radiography that a compression ratio of between 8 : 1 and 16 : 1 was visually lossless to most observers, and that at these ratios JPEG compression resulted in a performance that was similar to wavelet compression. The results from our study, however, demonstrated that wavelet-compressed images were better than JPEG images at the same compression ratio, which agrees with other medical studies.5 ± 10 We did not examine the performance of the two compression methods at other ratios. The compression level was chosen on the basis of the results of our previous study where JPEG 27 reduced
Table 5 The inter-observer reliability coefficient of the nine observers for the three modalities Modality JPEG Wavelet Original 300 d.p.i.
Reliability score 0.96 0.97 0.96
Figure 3 Average absolute observer error scores for nine observers for all modalities. Five out of nine performed at the same level
the ®le size from 234 kB to 16 kB, a compression ratio of 14 : 1, when applied to clinical bitewings from a direct digital system (Digora Storage phosphor plate, Soredex, Helsinki, Finland).14 As in the present study, the ®le size of the bitewing radiographs digitized with a ¯atbed scanner at 300 d.p.i. was 150 kB. Compressing with JPEG at level 27 resulted in a ®le size of 16 kB, or a compression ratio of 9 : 1. The dierence in compression ratio could be caused by a dierent noise level in the images included in both studies.14,20 The average SD of the gray values of the subtraction images of the original and the wavelet-compressed images was smaller than that of the original and JPEG images, which could indicate that the wavelet compression algorithm at the same compression level preserves more information in the image than JPEG. The dierence of the SD was not statistically signi®cant and the dierence may not be clinically signi®cant. Trapnell et al.15 examined the eect of wavelet compression (25 : 1, 50 : 1, 75 : 1 and 100 : 1) on the diagnostic accuracy of speci®c osseous changes in a series of digitized TMJ radiographs. They found that the diagnostic accuracy was not in¯uenced by the level of wavelet compression and that wavelet-compressed images provided increased diagnostic accuracy over ®lm in assessing the condylar position. In the present study, a slightly better performance was also found in wavelet-compressed images in detecting dentinal caries. This may have been a result of the `de-noise' eect of the compression algorithm which is also seen in other studies.7,14,15 This de-noise eect has been described as the eect of using a `low-pass ®lter'.4 In this study, a signi®cant dierence between wavelet and JPEG compression was found only for dentinal lesions. Although more research is required for a full explanation of this phenomenon, results from other studies give some indications. It is well known from perception theory that the contrast and contour de®nition of an object complement each other. The detectability of two objects of the same diameter, one with low contrast and well delineated and the other showing high contrast but with an indistinct contour, is comparable.21 In dental radiographs, dentinal lesions have a lower contrast than enamel lesions. It is also true that the radiographic outline of caries lesions is fairly indistinct. From the subtraction images (Figure 1) it can be seen that the JPEG algorithm tends to reduce the image contrast more than the wavelet technique does. While the wavelet subtraction image does not show a speci®c pattern that can be matched with image details, the JPEG subtraction image shows darker areas corresponding with the brighter areas in the original image. The fact that these areas in the subtraction image are darker, indicates that they are less bright in JPEG-compressed images, resulting in a lower contrast ratio of these areas and the background. Figure 4 is an enlargement of part of the images shown in Figure 1, but at higher JPEG and wavelet compression ratios and it can be seen that the JPEG algorithm results in block artifacts aecting the
261
Dentomaxillofacial Radiology
Compression algorithms A Janhom et al
262
Figure 4 Enlarged region of interest obtained from a JPEGcompressed image and from a wavelet-compressed image. The JPEG image shows block-like artifacts, while the artifact in the wavelet image follows the contours of image details more closely
continuity of contours. This is a direct consequence of the way in which JPEG reduces image information, by mapping the image into groups of 868 pixels. The wavelet technique produces artifacts as well, but these follow more or less the contours of image details. The compression ratio used in the examples presented in Figure 4 is higher (25 : 1) than the one used in our study, but the eects of JPEG and wavelet compression are likely to be the same at lower compression ratios, albeit less pronounced. Therefore, it is very plausible that the more unfavorable eect of JPEG compression both on contrast and on contour reproduction explains not only the lower diagnostic quality of JPEGcompressed images, but also the more signi®cant eect on low contrast, radiographically ill-de®ned dentinal lesions than on enamel-related radiolucencies. As far as we know, this is the ®rst wavelet compression study for caries detection. The compres-
sion level tested, which resulted in a ®le size reduction to 10% of the original size, is generally considered to be a low compression rate. Higher compression rates are possible. However, a previous JPEG compression study showed that JPEG 53, which resulted in a reduction to 5%, aected the diagnosis of enamel lesion unfavorably.20 The eect on caries diagnosis of higher compression levels using the wavelet method is yet to be examined. The wavelet implementation in dierent software programs can provide dierent results.7 Other studies have shown that some packages performed better than JPEG while others were similar to JPEG.4,6,10 When applied to various medical images the results were also mixed, some showing a better quality than JPEG and others showing similar outcomes. The variations may be a product of the dierent software packages used; other packages may give dierent results from those we found in our study. Further research will show if this is a consistent eect. In conclusion, this study shows that at a 9 : 1 compression ratio, JPEG and wavelet algorithms were equally good for the detection of initial approximal caries lesions and lesions up to the DEJ. When compared to the original digital image scanned at 300 d.p.i., wavelet compression produced no adverse eects for the diagnosis of caries lesions at any depth. The JPEG-compressed images performed signi®cantly worse than the original and wavelet-compressed images for the detection of dentinal lesions. It seems that in general wavelet compression is a better choice than JPEG at the same compression ratio.
Acknowledgements Our thanks are due to all nine observers who participated in this study, and to Dr J Peter van Amerongen for his help in the histology examination. The authors wish to thank Dr M Kevin O Carroll for editorial assistance.
References 1. Rabbani M, Jones PW. Digital image compression techniques. SPIE Tutorial texts series. Vol. TT7. Bellingham, WA: SPIE Optical Engineering Press. 1991. 2. Good WF, Maitz GS, Gur D. Joint Photographic Experts Group (JPEG) compatible data compression of mammograms. J Digit Imaging 1994; 7: 123 ± 132. 3. Cox GG, Cook LT, Insana MF, McFadden MA, Hall TJ, Harrison LA, et al. The eects of lossy compression on the detection of subtle pulmonary modules. Med Phys 1996; 23: 127 ± 132. 4. 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 imagecomparison workstation. Radiology 2000; 215: 543 ± 553. 5. Erickson BJ, Manduca A, Palisson P, Persons KR, Earnest F, Savcenko V, et al. Wavelet compression of medical images. Radiology 1998; 206: 599 ± 607.
Dentomaxillofacial Radiology
6. Ricke J, Maass P, Hanninen EL, Liebig T, Amthauer H, Stroszczynski C, et al. Wavelet versus JPEG (Joint Photographic Experts Group) and fractal compression. Impact on the detection of low-contrast details in computed radiographs. Invest Radiol 1998; 33: 456 ± 463. 7. Erickson BJ, Manducs A, Persons KR, Earnest IV F, Hartman TE, Harms GF, et al. Evaluation of irreversible compression of digitized posterior-anterior chest radiographs. J Digit Imaging 1997; 10: 97 ± 102. 8. Goldberg MA, Pivovarov M, Mayo-Smith WW, Bhalla MP, Blickman JG, Bramson RT, et al. Application of wavelet compression to digitized radiographs. Am J Radiol 1994; 163: 463 ± 468. 9. Schomer DF, Eleker AA, Hazle JD. Introduction to waveletbased compression of medical images. Radiographics 1998; 18: 469 ± 481.
Compression algorithms A Janhom et al
10. Iyriboz TA, Zukoski MJ, Hopper KD, Stagg PL. A comparison of wavelet and Joint Photographic Experts Group lossy compression methods applied to medical images. J Digit Imaging 1999; 12: 14 ± 17. 11. Van der Stelt PF, Sanderink GCH, Dula K, Huiskens R. Lossy ®le compression and diagnostic image quality of digital intraoral radiographic images. J Dent Res 1997; 76 (special issue): 140 (Abstr1010). 12. Sanderink GCH, Dula K, Huiskens R, Van der Stelt PF. The loss of image quality in digital panoramic radiography using image compression. In: Farman AG, Ruprecht A, Gibbs SJ, Scarfe WC, (eds). Advances in Maxillofacial Imaging. Amsterdam: Elsevier, 1997; pp 299 ± 305. 13. Wenzel A, Gotfredsen E, Borg E, GroÈndahl H-G. 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. 14. Janhom A, Van der Stelt PF, Van Ginkel FC, Geraets WGM. Eect of noise on the compressibility and diagnostic accuracy for caries detection of digital bitewing radiographs. Dentomaxillofac Radiol 1999; 28: 6 ± 12. 15. Trapnell CJ, Scarfe WC, Cook JH, Silveira AM, Regennitter FJ, Haskell BS. Diagnostic accuracy of ®lm-based, TIFF, and wavelet compression digital Temporomandibular joint images. J Digit Imaging 2000; 13: 38 ± 45.
16. Jones DEA, Raine HC, Mix D. Correspondence to the editor. Br J Radiol 1949; 22; 549 ± 550. 17. Jones DEA. A note on back-scatter and depth doses for elongated rectangular X-ray ®elds. Br J Radiol 1948; 22: 342 ± 345. 18. Svenson B, Peterson A. In¯uence of dierent developing solutions and developing times on radiographic caries diagnosis. Dentomaxillofac Radiol 1990; 19: 157 ± 160. 19. Janhom A, Van Ginkel FC, Van Amerongen JP, Van der Stelt PF. Scanning resolution and the detection of approximal caries. Dentomaxillofac Radiol 2001; 30: 166 ± 171. 20. Janhom A, Van der Stelt PF, Van Ginkel FC. Interaction between noise and ®le compression and its eect on the recognition of caries in digital imaging. Dentomaxillofac Radiol 2000; 29: 20 ± 27. 21. Geleijns J, Schultze Kool LJ, Zoetelief J, Zweers D, Broerse JJ. Image quality and dosimetric aspects of chest x-ray examinations: measurements with various types of phantoms. Proceeding of Radiation Protection Dosimetry 1993; 49: 83 ± 88.
263
Dentomaxillofacial Radiology
