Sparseness of the trabecular pattern on dental radiographs: visual assessment compared with semi-automated measurements

Abstract

Objective

In diagnostic imaging; human perception is the most prominent, yet least studied, source of error. A better understanding of image perception will help to improve diagnostic performance. This study focuses on the perception of coarseness of trabecular patterns on dental radiographs. Comparison of human vision with machine vision should yield knowledge on human perception.

Method

In a study on identifying osteoporotic patients, dental radiographs were made from 505 post-menopausal women aged 45–70 years. Intra-oral radiographs of the lower and upper jaws were made. Five observers graded the trabecular pattern as dense, sparse or mixed. The five gradings were combined into a single averaged observer score per jaw. The radiographs were scanned and a region of interest (ROI) was indicated on each. The ROIs were processed with image analysis software measuring 25 image features. Pearson correlation and multiple linear regression were used to compare the averaged observer score with the image features.

Results

14 image features correlated significantly with the observer judgement for both jaws. The strongest correlation was found for the average grey value in the ROI. Other features, describing that osteoporotic patients have fewer but bigger marrow spaces than controls, correlated less with the sparseness of the trabecular pattern than a rather crude measure for structure such as the average grey value.

Conclusion

Human perception of the sparseness of trabecular patterns is based more on average grey values of the ROI than on geometric details within the ROI.

Image perception is an important aspect of diagnostic imaging [1,2]. According to the UNSCEAR 2000 report (Volume 1: Annex D), the average number of diagnostic radiological examinations in countries with Level I healthcare is about 1000 per year per 1000 population. Therefore, it can be estimated that each European has about one radiological examination per year.

The interpretation of radiographs is complicated by the variations in human anatomy and the spatial information that is lost while projecting the patient body on a two-dimensional plane [3]. Visual clues are overlooked or misinterpreted [4-6]. The diagnostic process of radiologists can be improved by the use of computers [7-11]. Pattern-recognition techniques have been designed to draw the attention of the radiologist to regions in mammograms that need careful scrutiny and interpretation [12]. Fully automated methods can screen chest radiographs for features of tuberculosis [13]. Although the results compete with human performance, the automated methods do not outperform the radiologists. It is expected that computers may eventually replace human observers in the analysis of data. However, complete replacement of the human observer is still a remote possibility [11]. For the foreseeable future, human interpretation will continue to be an inseparable element of medical imaging [14]. We need to understand the images and the technologies used to acquire and display them, but as patient treatment and care depend to a large extent on radiologists interpreting images, we also need to understand human perception and cognition. Many parameters are involved in the process of image acquisition, image processing and image display, and it is largely unknown how they should be optimised for human interpretation. Understanding the perceptual and cognitive processes involved in reading medical images will help to enhance the most useful properties of the images to improve diagnostic performance and reduce error rates [2,3,6,14-17].

In dental radiography, many radiographs show bone with a radiographic trabecular pattern, an irregular meshwork of vague bright lines with fuzzy dark meshes (Figure 1). Visual assessment of the trabecular pattern in intra-oral radiographs is a method to identify women at risk of having osteoporosis. Dense trabeculation is a strong indicator of healthy bone, whereas sparse trabeculation is a sign of osteoporosis [18-20].

Figure 1.

Radiograph of the right side of the lower jaw with region of interest 3.7×5.8 mm between the first and second premolar. This is used to measure mean and standard deviation of the grey value.

At the oral radiology department of the Academic Centre for Dentistry Amsterdam, methods were developed for semi-automatic analysis of the trabecular pattern of radiographs. Measurements on the trabecular pattern of intra-oral radiographs were found to predict bone mineral density and osteoporosis [21,22].

When both the visual assessment and the semi-automatic analysis had been applied to the same set of radiographs, there arose an opportunity to compare the two and to gain more insight into the human perception of the coarseness of the radiographic trabecular pattern.

Methods and materials

In 2003, the European Union granted a research project, named OSTEODENT, to five European universities at Manchester, Amsterdam, Athens, Leuven and Malmö.

Subjects and radiographs

In the project, women from Manchester, Athens, Leuven and Malmö were recruited [20,23]. Local ethical approval for the study was obtained in each recruiting centre and informed consent was obtained from all subjects. From each subject, intra-oral radiographs were made from the upper right and lower right premolar region using one of three Planmeca Prostyle Intra devices (60–63 kV; Planmeca Oy, Helsinki, Finland) or with a Siemens Heliodent MD (60 kV; Sirona, Bensheim, Germany).

The radiographic trabecular pattern was graded by three experienced radiologists and two general practitioners [20]. They were given three reference images from the upper and the lower jaw, and they were asked to classify the trabecular pattern between the roots of the premolars as dense, alternating dense and sparse, or sparse. Additional instructions to observers described dense trabeculation as having many trabeculae connected to each other and small or few marrow spaces. Sparse trabeculation was described as having fewer trabeculae, larger marrow spaces and being darker. Any trabecular pattern that was ambiguous had to be assigned to the intermediate category. Lamina dura, mandibular cortex and maxillary sinus, as well as diseased areas, were excluded from the assessment.

Subjects who had not been graded by all the observers were excluded from the study [20]. Bone mineral density (BMD) values were measured by dual energy X-ray absorptiometry scans of the left hip and the lumbar spine. Complete sets of data, including BMD of hip and spine, two intra-oral radiographs and five observer gradings, were obtained from 505 subjects, of whom 21% were diagnosed as osteoporotic in accordance with the criteria of the World Health Organization if the T-score of the hip or spine was −2.5 or less.

Image processing

The intra-oral radiographs were scanned with a flatbed scanner (Agfa Duoscan T₁200; Agfa Gevaert, Mortsel, Belgium; fixed sensitivity settings) at a resolution of 118 pixels per centimetre (300 dpi). Most radiographs displayed three interdental regions, of which the widest was used by an observer to select a region of interest (ROI) containing a trabecular pattern only (Figure 1). The ROI was subjected to automated image analysis procedures measuring various image features that have proven their relevance for bone structure and osteoporosis extensively [21-27]. Firstly, the mean (MEAN) and standard deviation of the grey values were determined on the raw unfiltered ROI (Figure 1).

Isolated pixels with deviating grey values were adjusted with a 3×3 median filter. Large-scale variations in grey value caused by varying thickness of cortex and soft tissues were removed with an unsharp self-masking filter. Then the ROI was segmented using the mode of the histogram as the threshold value. This resulted in a version of the ROI consisting of black and white segments (Figure 2). The segments were used to measure the fractal dimension according to the caliper method (FRACT), the number of black segments (N_black), the number, area and the perimeter of the white segments (N_white, BV/TV, BS/TV, respectively) and an index of orientation in horizontal direction (from 0° to 165° in steps of 15°: LFD 0, LFD 15, …, LFD 165).

Figure 2.

The region of interest in Figure 1 has been filtered and segmented into black and white segments. This is used to measure fractal dimension, numbers of black and white segments, area and perimeter of the white segments and orientation.

Next, the white segments were eroded to a wire frame (Figure 3) that was used to measure the length of the frame, the number of terminal points and the number of nodes (TSL_white, N.Tm_white, N.Nd_white, respectively). Similarly, the black segments were eroded to a wire frame that was used to measure the length, the number of terminal points and the number of nodes (TSL_black, N.Tm_black, N.Nd_black, respectively). Various methods for filtering and measuring image features have been described before [21,23-32]. Similar image features are current in studies on osteoporosis and bone structure [33-39].

Figure 3.

The white segments in Figure 2 have been eroded. The eroded parts are displayed in grey. The remaining wire structure is displayed in white. Each white pixel contributes to the length of the white frame (TSL_white). Each white pixel with one (or no) white neighbours is an end point and contributes to the number of end points (N.Tm_white). Each white pixel with three (or four) white neighbours is a node and contributes to the number of nodes (N.Nd_white).

Statistics

To reduce the variations between the individual gradings and to simplify the analysis, the gradings of the five observers were combined by equating the gradings dense, alternating and sparse with numbers 1, 2 and 3, respectively. For each subject and each jaw the five numbers were combined into a single averaged observer score, resulting in 1010 averaged observer scores pertaining to 505 subjects.

For the interobserver agreement, values of the kappa index ranged from 0.32 (fair) to 0.55 (moderate) [20]. As kappa is the agreement, it can be seen that the disagreement, or noise, in the individual observer is 0.45–0.68. To estimate the noise in the averaged observer score, the factor 1/√5 is used, leading to a noise level of 0.20–0.30. Obviously 30% of the variation in the averaged observer judgement must be considered noise and 70% of the variation in the averaged observer score is the maximum that can be accounted for by any set of features.

With respect to the image features, it can be said that the semi-automatic measurements were very reproducible as most of the associated values of Cronbach's α exceeded 0.8, and even 0.9 [27].

To test the relation between the averaged observer score and the image features, the Pearson correlation was calculated. In addition, stepwise multiple linear regression was applied to calculate the multiple correlation between the averaged observer score and the image features. Single and multiple correlations were computed with the SPSS package (version 18; SPSS Inc., Chicago, IL). To define significance, α=0.05 was used. Additional computations of confidence intervals were done in accordance with Hays [40].

Results

Table 1 summarises the correlations of the image features on the lower and upper jaws with the averaged observer score. Of the 25 features that were investigated, 14 correlated significantly for both jaws and 6 correlated significantly for only 1 jaw. The difference between the correlation coefficients for the upper and lower jaws was less than the critical value of 0.124 for all features except BS/TV. Allowing 1 in 20 features to differ, this implies that the corrrelations for the upper and lower jaws correspond.

Table 1. Correlations between observer grading and image features. Osteoporotic patients tended to have high observer gradings, low grey values and few details. Healthy controls tended to have low observer gradings, high grey values and many details.

Feature	Code	Upper jaw	Lower jaw
Average grey value	MEAN	−0.37	−0.39
Standard deviation of grey value	SD	−0.10	−0.13
Fractal dimension	FRACT	+0.16	+0.09
Number of white segments	Nwhite	+0.15	+0.11
Number of black segments	Nblack	−0.18	−0.18
Area of white segments	BV/TV	−0.19	−0.26
Perimeter white segments	BS/TV	−0.16	−0.32
Orientation horizontal	LFD 0	ns	+0.12
Orientation along 15°	LFD 15	ns	ns
Orientation along 30°	LFD 30	ns	ns
Orientation along 45°	LFD 45	ns	−0.13
Orientation along 60°	LFD 60	−0.14	−0.18
Orientation along 75°	LFD 75	−0.15	−0.19
Orientation vertical	LFD 90	−0.13	−0.16
Orientation along 105°	LFD 105	−0.10	−0.15
Orientation along 120°	LFD 120	−0.11	ns
Orientation along 135°	LFD 135	ns	ns
Orientation along 150°	LFD 150	ns	+0.11
Orientation along 165°	LFD 165	ns	+0.11
Length white wire	TSLwhite	−0.21	−0.32
End points white wire	N.Tmwhite	ns	ns
Nodes white wire	N.Ndwhite	−0.14	−0.19
Length black wire	TSLblack	ns	ns
End points black wire	N.Tmblack	−0.19	−0.22
Nodes black wire	N.Ndblack	−0.09	ns

ns, not significant.

The highest correspondence is found for the average grey value (MEAN) in the lower jaw (−0.39), as well as in the upper jaw (−0.37). This implies that MEAN accounts for 15% of the variance in the averaged observer score in the lower jaw and 14% in the upper jaw.

Using stepwise multiple linear regression, the percentage variation accounted for n increased to 19% (r=0.43) for the upper jaw and 27% (r=0.52) for the lower jaw. The analysis started with predictor MEAN and then predictors were added one by one until the prediction improved insignificantly. Table 2 shows the predictors and the order in which they entered the regression equation. Six predictors were entered for the lower jaw and four for the upper jaw. The three most important features in lower and upper jaw corresponded; they were MEAN, BS/TV and LFD 75.

Table 2. Results of stepwise multiple linear regression.

Upper jaw
Order in regression	Feature in upper jaw	Code	Variance accounted for
1	Average grey value	MEAN	14%
2	Perimeter white segments	BS/TV	16%
3	Orientation along 75°	LFD 75	18%
4	Standard deviation of grey value	SD	19%

Lower jaw
Order in regression	Feature in lower jaw	Code	Variance accounted for
1	Average grey value	MEAN	15%
2	Perimeter white segments	BS/TV	21%
3	Orientation along 75°	LFD 75	23%
4	Length of white wire	TSLwhite	25%
5	Orientation horizontal	LFD 0	26%
6	Nodes white wire	N.Ndwhite	27%

Stepwise multiple linear regression first selects the single feature best describing the observer grading. The second feature is selected to increase the descriptive power the most. This is continued as long as significant improvement can be achieved.

Discussion

The assessments of the trabecular pattern by the observers were similar for the upper and lower jaw [20]. In addition, the semi-automatic measurements of upper and lower jaws correspond to a large extent [29]. This may explain the correspondence of upper and lower jaws in Table 1.

In a previous study, several features discriminated significantly between osteoporotic patients and non-osteoporotic controls, of which the strongest was the N.Tm_black [29]. To a large extent this is confirmed by the correlations provided in Table 1. It can be seen that the trabecular patterns of osteoporotic patients have less geometrical detail than the patterns of the non-osteoporotic controls. Obviously, osteoporotic patients have fewer but bigger marrow spaces than non-osteoporotic controls. This is consistent with the finding that sparse trabeculation is a sign of osteoporosis.

The negative value of the correlation between MEAN and the observer score implies that low values of MEAN are associated with sparse trabecular patterns, which is consistent with the loss of bone mineral and increased sparseness of the radiographic trabecular pattern of osteoporotic patients [20,41]. Considering that MEAN represents the average grey value of the pixels in the ROI, one might infer that MEAN is a crude measure for structure. Therefore, it is striking that in this study MEAN has a stronger correlation with the observer grading than the other features (such as N.Nd_blackand TSL_white) that reflect relevant structural aspects of trabecular microarchitecture. MEAN even surpasses the features N_blackand BV/TBV, which are closely related with the concept of sparse and dense trabeculation. Part of a possible explanation can be found in the instructions to the observers stating that sparse trabeculation associates with darker images.

Our main conclusion is that the human perception of the sparseness of trabecular patterns is based more on the average grey value of the ROI than on other structural aspects of the ROI.