Diagnostic accuracy of direct digital dental radiography for the detection of periapical bone lesions

Diagnostic accuracy of direct digital dental radiography for the detection of periapical bone lesions II. Effects on diagnostic accuracy after application of image processing Boel Kullendorff, DDS, a and Mats Nilsson, PhD, b M a l m t , Sweden CENTRE FOR ORAL HEALTH SCIENCES AND MALMO UNIVERSITY HOSPITAL, LUND UNIVERSITY

Objectives. The diagnostic accuracy of direct digital radiography for the detection of small experimentally made periapical lesions was evaluated. A comparison was made between original digital images and images processed with different enhancement facilities of the digital system. Study design. Seven observers assessed the digital images in original mode and after individual image treatment. The diagnostic accuracy was calculated. The processing functions used by the observers were recorded, and the effect of processing was evaluated. Results. The overall diagnostic accuracy was not different for the two image modes. Neither were the individual results of the observers different. The image processing was most effective when alterings of contrast and brightness were used. More complicated processing procedures had less effect on the diagnostic accuracy. Conclusions. Image processing of direct digital images of high quality has a limited effect on the diagnostic accuracy. Basic processing functions, that is, alterings of contrast and brightness, were preferred for the detection of periapical lesions. (Oral Surg Oral Med Oral Pathol Oral Radiol Endod1996;82:585-9)

The occurrence of pathologic changes associated with periapical bone tissue is common, and radiography is still the superior tool for the detection of periapical pathosis. However, radiography is not a perfect diagnostic test. A lot of effort has been made to improve the imaging of the periapical bone in order to detect early signs of pathosis or healing. The development of digital radiography has created new options in this field. One of the first digital methods to be investigated was subtraction. 1 This technique increased the diagnostic accuracy in the detection of small periapical density changes as compared with conventional radiography. 2 However, the subtraction technique is difficult to use because an almost perfect geometric conformity is required between the images to be subtracted. An alternative of extracting more information is to use various kinds of image processing on the single digital image in order to allow the enhancement of special features. Over the last few years, several commercial systems for direct digital radiography (DDR) have been introduced. One of the advantages of this method compared with conventional radiography (based on analogue images) is the possibility of performing image processing or more exactly postprocessing. This Grants were received from the Swedish Dental Society and the Faculty of Odontology, Lund University. aDepartment of Oral Radiology, Centre of Oral Health Sciences. bDepartment of Radiation Physics, Malmt~ University Hospital, Lund University. Received for publication Feb. 14, 1996; accepted for publication May 10, 1996. Copyright 9 1996 by Mosby-Year Book, Inc. 1079-2104/96/$5.00 + 0 7/16/76147

improves the visual appearance of an image to the human eye and prepares the images so that features and structures present in the image can be measured. In a study of dentists' perception of the quality of digital radiographs, Wenzel and Hintze 3 found that most denfists preferred a treated image to the original one. Furthermore, image treatment increased the accuracy of caries diagnosis and the detectability of bone lesions, particularly in low density images. 4-6 For the measurement of intraosseous lesion dimensions, gray-scale equalized images resulted in better reliability than unenhanced or contrast-stretched images. 7 The aims of this study were (1) to compare the overall diagnostic performance of DDR for the detectability of periradicular lesions of the original digitally captured images with those treated with different image-processing facilities, (2) to examine how dentists used the image-processing facilities to enhance the digital image, and (3) to examine how the diagnostic performance was influenced by image processing.

MATERIAL AND METHODS X-ray source and DDR system The x-ray source (Siemens Heliodent, Siemens, Erlangen, Germany) was operated at 70 kVp, 10 mA, and a total filtration of 2 m m AI equivalent. The distance between focus and collimator opening was 20 cm. The Visualix/VIXA system (Gendex Dental Systems, Milano, Italy) was equipped with a chargecoupled device (CCD) sensor of 18.1 x 24.2 m m active area. The image matrix was 288 x 384 with a pixel size of 63 x 63 tam. 8 The high-contrast resolution limit of this system has been calculated to 7 to 585

586

Kullendorffand Nilsson

ORAL SURGERY ORAL MEDICINE ORAL PATHOLOGY

November 1996 l a n e I. Observer ROC area values expressed as P(A) and Az for the diagnostic accuracy in detecting experimentally made periapical lesions Area P(A) observer 1 2 3 4 5 6 7 Mean P(A) Mean Az

DDR original i m a g e 0.80 0.70 0.74 0.79 0.87 0.69 0.69 0.75 0.79

DDR processedimage 0.83 0.64 0.71 0.78 0.84 0.77 0.71 0.75 0.79

8 lp/mm. 9' 10 Images were displayed on an Ultra V G A monitor (Lite-on, Taipei, Taiwan) operated with a screen resolution of 1024 x 768 pixels.

Periapical bone lesions Six dry human mandibles, free from periapical pathosis, were sectioned and divided into one buccal and one lingual part. The teeth were glued to the lingual part. With the buccal part removed, stepwise extended lesions were made at the apex of one molar or one premolar of each specimen that involved the lamina dura. The lesions were made with dental burs, 1 to 5 m m in diameter. Before x-ray exposures were made, the buccal bone plate was replaced. The corresponding teeth of the contralateral mandibles served as lesion-free specimens. The technical procedures are described in previous studies. 2, 9 In all, 20 images of roots with a lesion and 16 without a lesion were produced.

Examination with direct digital radiography The bone specimens were mounted in a silicon paste block and placed in a water-filled plexiglass box. The sensor was placed in a holder on the outside of the box. The x-ray tube was placed on the other side of the box with the collimator opening parallel to the sensor plane. Images of each lesion step were produced with the Visualix/VIXA system. The images were stored in the computer in their original form; contrast and brightness were set by the exposure factors. The monitor displayed the images on a 6-bit gray scale, corresponding to 64 gray levels. A reference step-wedge image was used to calibrate contrast and brightness settings of the screen.

Interpretation of images Seven observers, four oral radiologists and three endodontists, participated in this study. The digital images were displayed on the CRT monitor of the Visualix/VIXA system. The readings were made with standardized viewing conditions of subdued light. The observer to screen distance was about 60 cm. The images were presented in random order with respect to

Table II. Distribution of diagnostic decisions of seven observers after application of image processing compared with their decisions for the 36 original digital images. Diagnostic outcome was classified as improvement, no change, or impairment. No processing was performed on 16 readings. No Observer hnprovement change 1 2 3 4 5 6 7 Total

8 3 9 2 3 15 5 45

25 20 23 30 24 15 29 166

Number of images Impairment processed 1 6 4 3 3 6 2 25

34 29 36 35 30 36 36 236

presence or absence of lesions. The observers were informed about which root to examine and that the same tooth would appear in more than one image, representing different periapical status. They were asked to indicate on a 5-point rating scale the certainty with which a lesion was present or not present for each root. In an earlier separate session, the observers had been informed about the possibilities of image processing. They were also presented two test images, one with and one without a lesion, and given the option of trying the various processing procedures. Initially, the original digital image was shown with default brightness and contrast settings by the computer system. The observers made their first diagnostic decision on the basis of this image. After this they were allowed to process the image in the way they preferred in order to obtain what they found to be the best possible presentation, that is, a subjective evaluation of image quality. A second diagnostic decision was made after this. One of us (B.K.) was present during all observations, and the processing chosen for each image was recorded. Alternative choices for the same image were also recorded. The observers had no time limit. The following image-processing functions, mainly as described in the reference manual of Gendex Visualix, 11 were presented to the observers: gray-scale treatment functions and filtering treatment functions.

Gray-scale treatment functions. Equalization (E) activates an algorithm for gray-scale equalization. The gray level spectrum of the image is stretched out to fully use the entire gray-scale range of the system. This is also known as histogram equalization. Actual improvement in image quality may vary from image to image. Less level~more level decreases or increases by a factor 2 the number of displayed gray levels. The m a x i m u m and default number is 64. Contrast (C+, C-), the increase or decrease of image contrast by

Kullendorffand Nilsson 587

ORAL SURGERY ORAL MEDICINE ORAL PATHOLOGY

Volume 82, Number 5

Fig. 1. A, Original digital image of mandibular premolar with small experimentally made periapical lesion. B, Same image as A processed with decrease of brightness and increase of contrast. This resulted in improved diagnostic decisions for five of seven observers. C, Same image as in Fig. 2A processed with the equalizing function; resulted in improved diagnostic decision for one observer. fixed steps, is related to the slope of the density diagram of points in the image versus the dose imparted to the sensor at that point. Brightness (B+, B-) is the increase or decrease in image brightness by shifting the video output level in fixed steps of 10%. Filtering treatment functions. Sharpening (S) applies a band pass filter at high frequencies (simultaneously an edge enhancement and a mild noise removal filter). The resulting image is sharper. Edge enhancement (H) is a spatial differentiation filter in which a subtraction is made between the image and the image that has been shifted obliquely a few pixels. In this way a kind of image gradient is obtained that shows the edges of the objects. Smoothing (M) applies a smoothing spatial filter to remove image noise. The image becomes slightly blurred. The Visualix system is also equipped with a zooming function, but this function was not used in this study.

Analysis of the observations The observers' answers for each image were compared with the results from the assessment of the original image. All alternative processing was recorded for each image. However, only one diagnostic decision was set for each processed image. The different processing functions used were described as being either single or combined functions. Processing functions used in only a few cases, were pooled as "other functions." As all alternatives were recorded, the total number of readings on processed images was increased from 236 to 293. The readings were classified as improvement of diagnostic decision, no Change, or impairment. The diagnostic accuracy for each ob-

Table Ul. Distribution of image treatment functions and their effect on the diagnostic decisions. Treatment functions were: increase or decrease of brightness (B+ and B-), increase of contrast (C+), equalization (E), sharpening (S), smoothing (M), and other functions. Effects on the diagnostic decision were classified as improvement, no change, or impairment. All image treatment alternatives were counted resulting in 293 readings. Processing function B - C+ C+ E B+ C+ S and M B - C+ and S Other functions Total

No. of images Impairmentprocessed

Improvement 16 7 10 6 3 2 9 53

52 53 33 13 14 t2 33 210

8 7 4 1 2 4 4 30

76 67 47 20 19 18 46 293

server and image mode, as well as for all observers, was calculated as the areas under receiver operating characteristic (ROC) curves, the P(A) value. The overall results were also calculated as the Az value using the " R O C F I T " program) 2 The mean values were statistically compared with the Wilcoxon's signed rank test.

RESULTS Diagnostic accuracy There was no difference between the two image modes. The overall diagnostic accuracy measured as P(A) value was 0.75 and as Az value 0.79 for both

588

ORAL SURGERY ORAL MEDICINE ORAL PATHOLOGY November I996

Kullendorffand Nilsson

Visible contrast (%) 100 50

//

20 10 5 2 1 0.5_

J

0.2 0.1

I

0.1

/

0.2

[

I

0.5

I

I

I

I

1

I

2

I

I

5

I

I

I

10

20

50

100

Spatial frequency (Ip/degree) Fig. 2. Contrast-detail curve for eye (modified from S.M Pizer, 1985) illustrates variance in spatial frequency of human eye for different degrees of contrast. image modes. The individual P(A) values of the original and processed images for seven observers showed no statistical difference (p < 0.2). These results are presented in Table I.

Image processing Table II shows the distribution of diagnostic decisions after application of image processing for the seven observers on the original 36 images. A representative original image of a premolar with a periapical lesion is shown in Fig. 1A. In 16 cases, 6% of all readings, no image processing was performed. For the processed images, the diagnostic decision was improved in 18% of cases, not changed in 66%, and impaired in 10% of all readings. Observers 1, 3, 6, and 7 improved the accuracy with image processing; observer 5 presented equal results, and observers 2 and 4 decreased the accuracy. Table IB presents the observers' choice of image-processing functions. The functions most frequently used were increase of contrast and decrease of brightness (B to C+), as demonstrated in Fig. lB. An increase of contrast as a single function and equalization, as demonstrated in Fig. I C, were also applied. Two observers, 2 and 3, used combinations of several processing functions.

DISCUSSION The diagnostic accuracy measured as overall and individual P(A) and Az values for the original and processed images was not different. However, when analyzing the diagnostic outcome of image processing, there was an improvement in 18% of the images.

Even if an impairment was obtained in 10% of the images, a positive effect of image treatment could be expected. One reason for this discrepancy can be the construction of the ROC curves, where a t w o - s t e p change on the rating scale has more of an influence than a one-step change. This will not be evident from the table, where only " c h a n g e " or " n o change" is presented. It can also be of importance at which one of the cutting points on the ROC curve the change was made. An increased rate of true-positive answers for the processed images can be neutralized by an increase in the false-positive rate. Such results have been shown for other digital systems. I3,14 The most effective processing functions overall were altering the level of brightness in combination with increasing contrast. Five of the seven observers mainly used the contrast/brightness functions or the equalizing function; two observers also used combinations of other functions. One of them (observer 2) often chose an alternative combination of increased contrast, decreased brightness, and sharpening filter. The other one (observer 3)preferred a combination of sharpening and smoothing filters as an alternative to alterings of contrast or brightness. None of these combinations was effective. The observer who achieved the most improvements (observer 6) mainly used the equalizing function. Apparently, the combination of the more sophisticated processing functions were of limited value in improving the diagnostic accuracy with respect to periapical lesions. This has also been found for the detection of cariesJ 5 The preferred

ORAL SURGERY ORAL MEDICINE ORAL PATHOLOGY Volume 82, Number 5 p r o c e s s i n g is likely task dependent, 3 and the possibility o f altering the i m a g e for different p u r p o s e s is an a d v a n t a g e o f the digital systems. H o w e v e r , these systems seems to b e p r o v i d e d with m o r e functions than are neceSsary for o r d i n a r y diagnostic purposes. Future b u y e r s o f D D R systems w o u l d benefit b y choosing between image-processing programs of v a r i o u s levels o f sophistication and cost and selecting those that best suit their i n d i v i d u a l needs. W e d e c i d e d not to use the z o o m i n g function o f the D D R s y s t e m as this w o u l d h a v e m a d e the evaluation o f p r o c e s s i n g effects m o r e c o m p l i c a t e d . T h e effects o f z o o m i n g h a v e recently b e e n investigated b y M r y s t a d et a1.16 w h o f o u n d that i m a g e m a g n i f i c a t i o n h a d an upper limit b e y o n d w h i c h diagnostic a c c u r a c y was reduced. In our studies, the default size o f the d i s p l a y e d i m a g e on the m o n i t o r was 95 x 65 mm. T h e pixels in the 384 x 2 8 8 m a t r i x are c o n s e q u e n t l y 0.25 x 0.25 m m . T h e i m a g e was v i e w e d at a distance o f about 60 cm. T h e C R T i m a g e will therefore d i s p l a y 21 lp/degree, and the e y e will r e s o l v e pixels as squares if the contrast is greater than 20%. H o w e v e r , the structures o b s e r v e d in the p e r i a p i c a l b o n e tissue h a v e far less contrast than 20%, w h i c h m a k e s a l o n g e r v i e w i n g distance unnecessary. W h e n v i e w i n g digital i m a g e s on a C R T monitor, a sufficient v i e w i n g distance is important. 17 T h e h u m a n e y e has a l i m i t e d spatial resolution, w h i c h d e p e n d s on contrast level and i m a g e brightness. F o r a t y p i c a l C R T l u m i n a n c e level, the resolution limit is around 30 l p / d e g r e e (line pairs p e r o p e n i n g angle as seen f r o m the eye) for 100% contrast ( b l a c k and white), 15 l p / d e g r e e for 5% contrast and 5 l p / d e g r e e for 1% contrast. T h e r e f o r e if a b l a c k and white line b a r pattern has m o r e than 30 lp/degree, the e y e will not b e able to separate the lines (Fig. 2). A digital i m a g e d i s p l a y e d on a C R T m o n i t o r and v i e w e d at a distance resulting in pixels larger than this criterion (for e x a m p l e 1/30 d e g r e e for 5% contrast) will seriously i m p a i r the o b s e r v e r ' s ability to find areas with low contrast. In this case, the o b s e r v e r actually sees the s q u a r e - s h a p e d pixels and interprets the i m a g e as a pattern o f squares rather than as an a n a t o m i c image. T h e faculty o f vision instinctively tries to find g e o m e t r i c figures in an image. Thus the v i e w i n g distance has to b e long e n o u g h so that the pixels in the i m a g e are s m a l l e r than the resolution limit for the eye. In this study, the original digital i m a g e s were all of high quality. A strength o f the digital systems is the p o s s i b i l i t y o f i m p r o v i n g an existing i m a g e o f low quality instead o f m a k i n g a s e c o n d exposure. F o r rad i o g r a p h s with i m p a i r e d density, it has b e e n shown that i m a g e e n h a n c e m e n t can significantly i m p r o v e the diagnostic accuracy o f e x p e r i m e n t a l l y m a d e cortical b o n e lesions in p i g mandibles. 4 T h e r e f o r e it m i g h t b e w o r t h w h i l e to e x a m i n e the diagnostic accu-

Kullendorffand Nilsson

509

racy o f p e r i a p i c a l lesions in digital i m a g e s o f various quality as is the case in clinical work. We are grateful to the seven observers for their participation and for valuable comments. REFERENCES

1. GrOndahl H-G, Grrndahl K, Webber RL. A digital subtraction technique for dental radiography. Oral Surg Oral Med Oral Pathol 1983;55:96-102. 2. Kullendorff B, Grrndahl K, Rohlin M, Henrikson C-O. Subtraction radiography for the diagnosis of periapical bone lesions. Endod Dent Traumatol 1988;4:253-9. 3. Wenzel A, Hintze H. Perception of image quality in direct digital radiography after application of various image treatment filters for detectability of dental disease. Dentomaxillofac Radiol 1993;22:131-4. 4. Wenzel A. Effect of image enhancement for detectability of bone lesions in digitized intraoral radiographs. Scand J Den Res 1988;96:149-60. 5. Wenzel A, Fejerskov O, Kidd E, Joyston-Bechal S, Groeneveld A. Depth of occlusal caries assessed clinically by conventional film radiographs and by digitized processed radiographs. Caries Res 1990;24:327-33. 6. Wenzel A, Fejerskov O. Validity of diagnosis of questionable caries lesions in occlusal surfaces of extracted third molars. Caries Res 1992;26:188-94. 7. Farman AG, Avant SL, Scarfe WC, Green DB. In vivo comparison of the VIXA-2 and Ektaspeed plus radiographs for the assessment of periapical lesion dimensions. Abstracts, 5th European Congress on Dental and Maxillo-Facial Radiology,Cologne, German. 1995. 8. Molteni R. Direct digital dental x-ray imaging with Visualix/ VIXA. Oral Surg Oral Med Oral Pathol 1993;76:235-43. 9. Kullendorff B, Rohlin M, Nilsson M. Diagnostic accuracy of direct digital dental radiography for the detection of periapical bone lesions. I. Overall comparison between conventional and direct digital radiography. Oral Surg Oral Med Oral Pathol Oral Radiol Endod In press. 10. Welander U, McDavid WD, Sanderink GCH, Tronje G, Mrmer A-C, Dove SB. Resolution as defined by line spread and modulation transfer functions for four digital intraoral radiographic systems. Oral Surg Oral Med Oral Pathol 1994;78:109-15. 11. VIXDB reference manual. Milano, Italy:Gendex Dental System, 1994. 12. Metz C. ROCFIT program. Chicago, Ill. 13. Yokota ET, Miles DA, Newton CW, Brown Jr CE. Interpretation of periapical lesions using RadioVisioGraphy. J Endod 1994;20:490-4. 14. Luostarinen T, Tammisalo T, Vi~h~italo K, Tammisalo E. Comparison of intraoral digital and film radiography for diagnosis of periapical bone lesions. Dentomaxillofac Radiol 1995;24:92-3. 15. Wenzel A, Hintze H, Mikkelsen L, Mouyen F. Radiographic detection of occlusal caries in noncavitated teeth: a comparison between conventional film radiographs, digitized film radiographs and RadioVisioGraphy. Oral Surg Oral Med Oral Pathol 1991;72:621-6. 16. MOystad A, Svanaes DB, Larheim TA, Grrndahl H-G. Effect of image magnification of digitized bite-wing radiographs on approximal caries detection: an in vitro study. Dentomaxillofac Radiol 1995;24:255-9. 17. Pizer SM. Psychovisual issues in the display of medical images. Proceedings NATO, Advanced Studies Institute on Pictorial Information Systems in Medicine, Chapel Hill, N,C., 1985. Reprint requests: Boel Kallendorff, DDS Department of Oral Radiology Centre for Oral Health Sciences Carl Gustavs v~ig34, S-214 21 MalmS, Sweden