Image analysis of microslalograms of the mouse parotid gland using digital
• • image processing K. Yosbiura, M. Ohki and N. Yamada
Department of Oral Radiology, Nagasaki University School of Dentistry, Nagasaki, Japan
Received 12 February 1990 and in final form 6 May 1990 We compared two digital-image feature-extraction methods for the analysis of microsialograms of the mouse parotid gland following either overfilling, experimentally induced acute sialoadenitis or irradiation. Microsialograms were digitized using a drumscanning microdensitometer. The grey levels were then partitioned into four bands representing soft tissue, peripheral minor, middle-sized and major ducts, and run-length and histogram analysis of the digital images performed. Serial analysis of microsialograms during progressive filling showed that both methods depicted the structural characteristics of the ducts at each grey level. However, in the experimental groups, run-length analysis showed slight changes in the peripheral duct system more clearly. This method was therefore considered more effective than histogram analysis. Keywords: Parotid gland; mice; radiography; digital signal processing
A critical problem in radiological diagnosis is both the lack of quantitative evaluation and the wide observer variation 1-5, the latter leading to a reduction in the overall accuracy. Several attempts have been made to analyse radiographic patterns, such as the structure of bone trabeculae'r'", quantitatively by means of either Fourier'"? or microdensitometric analysis 7 •8 • 1O• However, a fully effective method has yet to be discovered. Ericson 11 demonstrated a correlation between sialographic images and gland function but stressed that a means of quantitating the data was needed. The first such evaluation was performed not with the duct system but with gland size!", This quantitative analysis was subsequently used in the differential diagnosis of salivary gland tumours and post-irradiation atrophy':'. The duct system in a sialogram is, however, more difficult to describe objectively due to its complexity!". Digital image processing is agreed to be the best modality for quantitative analysis because of the feasibility it offers of performing a range of image analyses. Feature extraction of microsialograms of the isolated mouse parotid gland has been described by Yoshiura and Kanda 15 who showed that the grey-scale histogram reflected characteristic structural features consistent with the underlying pathology. On the other hand, pattern recognition from this type of image analysis proved more difficult. Run-length analysis, a texture analysis method!", has been successfully used for flood mapping of rivers in the USA 17. In the present study we have applied this technique to the branching structure of the duct system of the parotid gland in an attempt to resolve the process of pattern recognition.
Materials and methods Animals
The present study was based on a total of 50 mice. Data for 40 were taken from a previous study l 5 and an additional 10 mice used for the overfilling and postirradiation phases of the present investigation. Experimental procedures
Acute sialoadenitis was induced by injecting 3.0,u1 of normal saline into the right parotid glands of 40 animals. The animals were sacrificed over a period 60 days, as described previously'", The left glands served as the control. The right parotid glands of four animals were injected with an excess amount of contrast medium (4.0,a1)18 (Barexmolt S-100, Toho Chemical Co., Tokyo) while an optimum amount (2.0,u1) in the left served as controls. Six animals were irradiated with 60 Gy to the right parotid region: the left served as the control. The radiation fields were 10 x 20 mrrr' (Figure 1) and the radiation data 200kVp, 20mA, 1390s with 0.3mm copper and 1.0 mm aluminium filters and a FSD of 26cm. Microsialography was performed 90 days later. Since the irradiated glands had marked atrophy, the injected volume was reduced to 1.5,u1. Microsialography and histolo~~ were as described previously on the excised glands . Six normal and 19 inflamed glands were randomly selected from the earlier data 15. Of the inflamed glands, eight were in the initial, five the mid and six the recovery stages 18.
©
1991 Butterworth-Heinemann for IADMFR 0250-832 x 91/010017-08
Dentomaxillofac. Radiol., 1991, Vol. 20, February
17
Digital image processing in microsialography: K. Yoshiura et al. microsialograms were digitized by a drum-scanning microdensitometer as 512 x 512 x 8-bit images. Both sampling pitch and aperture size were 50,um. Images were transferred to a microcomputer (MicroVax-Il. DEC, USA) for analysis. The grey levels of all the digitized images were corrected by means of reference stepwedges and the densities equalized using the second polynomial equation by least squares approximation. In the digitized images, grey levels 0 and 255 represented complete black and complete white respectively; the intermediate densities were divided into 254 equal levels. Four cut-off points were selected by comparing all the microsialograms of the control group subjectively and set at 16, 32, 80, 128. Each image was partitioned on this basis into five revised grey-level bands: revised grey 0 (original grey 0-15), revised grey 1 (16-31), revised grey 2 (32-79), revised grey 3 (80-
5mm
20mm
..... c
o u
Q)
:0 e o
u
I
IOmm
l/l
3mm
Figure I Diagram showing the radiation field for the right parotid region: all areas of the gland are covered. P. parotid gland
Together with the four overfilled and six irradiated glands, a total of 35 glands were used for the analyses. Between three and six serial microsialograms were obtained of each gland.
a II
Digital image processing Figure 2 summarizes the image analysis procedure. The
s:
10 9 8
g.7 ~
6
c ~ 5
4
3 2
b'
- - - - - - - - - - - - 1- - - I- - - - - - - I- I- - I- - - I 1- - I I-
-
,-
4 4
I
2' Z Z
-
I-
I
-
-
- - -
I-
- - - - - - I- - - - .- - I -
I
I
I 'I I 'i - - ,- - i -
- - 'i T 2 -
l-I-
-
I-
12 10 >u
8
:> 13
6
eQ)
...
Q)
u.
4 2
Five-level quantitizatian
0
C Run length analysis
Figure 2 Flow diagram showing the sequence of image processing
18
Dentomaxillofac. Radiol., 1991, Vol. 20, February
2
3
4
5
6
Run length
Figure 3 Diagram to show run-length analysis. a. Binary figure simulating the duct. b. Run-length matrix. c. Histogram of the run lengths
Digital image processing in microsialography: K. Yoshiura et al. 127) and revised grey 4 (128-255). Revised grey 0 and 4 indicate complete black and white. These five bands defined the background, soft tissue, peripheral minor, middle-sized and major ducts on the microsialograms. Each image was then assessed for the run length of each of these grey levels which was displayed as the run-length histogram. The run length is the number of juxtaposed picture elements having the same grey level in one scan direction 16. For example, the run-length histogram of the binary figure shown in Figure 3a can be obtained by the following procedure: starting with the left-hand column, there are four fragments with a run length of 1. Moving to the next column on the right, there are again four fragments of one run length, and in the next two of two each. The overall matrix of the run length is shown in Figure 3b and the resulting histogram in Figure 3c. Run length was measured by scanning the image vertically in relation to the main excretory duct. The mean value and standard deviation of each value in the histogram was calculated together with the long and short runs emphasis!". Of these four parameters, long runs emphasis demonstrated slight differences in the duct pattern most clearly and was therefore selected as representative of the run-length histogram. It was calculated using the following equation:
where I = run-length value
F(/)
= frequency
of each run length.
The long runs emphasis was calculated for each revised grey level and also compared with the greyscale histogram. Background (grey level 0) was not analysed. Student's t test was used for the statistical evaluation.
Results Relationship between microsialographic appearance, amount of the contrast medium infused, grey-scale histogram and run-length analysis Typical control group microsialograms and 5-grey-level images are shown in Figure 4 and the related grey-scale histogram and run-length analysis in Figure 5. The grey-scale histogram shows that there are significant changes at grey level 3, and, to a larger extent, levels 1 and 4, following the injection of 0.5 and 1.0,u1. The run-length analyses, on the other hand, failed to demonstrate a similar increase in long runs emphasis at grey levels 3 and 4. At the latter, the only significant difference was on injection of 0.5,u1 and the former, with both 0.5 and 1.0,u1. In contrast, significant changes were observed at grey level 2.
Figure 4 Typical microsialograms (above) and their 5-level images (below) of a control parotid gland. The volumes of contrast medium injected are, from left to right, 0.5, 1.0, 1.5 and 2.0pl
Dentomaxillofac. Radiol., 1991, Vol. 20, February
19
Digital image processing in microsialography: K. Yoshiura et al. 15000
500
400 10000
** * •
..f/-\-\
(/)
'iii
(/)
0
s: 300 0-
Q)
>
• • image processing K. Yosbiura, M. Ohki and N. Yamada
Department of Oral Radiology, Nagasaki University School of Dentistry, Nagasaki, Japan
Received 12 February 1990 and in final form 6 May 1990 We compared two digital-image feature-extraction methods for the analysis of microsialograms of the mouse parotid gland following either overfilling, experimentally induced acute sialoadenitis or irradiation. Microsialograms were digitized using a drumscanning microdensitometer. The grey levels were then partitioned into four bands representing soft tissue, peripheral minor, middle-sized and major ducts, and run-length and histogram analysis of the digital images performed. Serial analysis of microsialograms during progressive filling showed that both methods depicted the structural characteristics of the ducts at each grey level. However, in the experimental groups, run-length analysis showed slight changes in the peripheral duct system more clearly. This method was therefore considered more effective than histogram analysis. Keywords: Parotid gland; mice; radiography; digital signal processing
A critical problem in radiological diagnosis is both the lack of quantitative evaluation and the wide observer variation 1-5, the latter leading to a reduction in the overall accuracy. Several attempts have been made to analyse radiographic patterns, such as the structure of bone trabeculae'r'", quantitatively by means of either Fourier'"? or microdensitometric analysis 7 •8 • 1O• However, a fully effective method has yet to be discovered. Ericson 11 demonstrated a correlation between sialographic images and gland function but stressed that a means of quantitating the data was needed. The first such evaluation was performed not with the duct system but with gland size!", This quantitative analysis was subsequently used in the differential diagnosis of salivary gland tumours and post-irradiation atrophy':'. The duct system in a sialogram is, however, more difficult to describe objectively due to its complexity!". Digital image processing is agreed to be the best modality for quantitative analysis because of the feasibility it offers of performing a range of image analyses. Feature extraction of microsialograms of the isolated mouse parotid gland has been described by Yoshiura and Kanda 15 who showed that the grey-scale histogram reflected characteristic structural features consistent with the underlying pathology. On the other hand, pattern recognition from this type of image analysis proved more difficult. Run-length analysis, a texture analysis method!", has been successfully used for flood mapping of rivers in the USA 17. In the present study we have applied this technique to the branching structure of the duct system of the parotid gland in an attempt to resolve the process of pattern recognition.
Materials and methods Animals
The present study was based on a total of 50 mice. Data for 40 were taken from a previous study l 5 and an additional 10 mice used for the overfilling and postirradiation phases of the present investigation. Experimental procedures
Acute sialoadenitis was induced by injecting 3.0,u1 of normal saline into the right parotid glands of 40 animals. The animals were sacrificed over a period 60 days, as described previously'", The left glands served as the control. The right parotid glands of four animals were injected with an excess amount of contrast medium (4.0,a1)18 (Barexmolt S-100, Toho Chemical Co., Tokyo) while an optimum amount (2.0,u1) in the left served as controls. Six animals were irradiated with 60 Gy to the right parotid region: the left served as the control. The radiation fields were 10 x 20 mrrr' (Figure 1) and the radiation data 200kVp, 20mA, 1390s with 0.3mm copper and 1.0 mm aluminium filters and a FSD of 26cm. Microsialography was performed 90 days later. Since the irradiated glands had marked atrophy, the injected volume was reduced to 1.5,u1. Microsialography and histolo~~ were as described previously on the excised glands . Six normal and 19 inflamed glands were randomly selected from the earlier data 15. Of the inflamed glands, eight were in the initial, five the mid and six the recovery stages 18.
©
1991 Butterworth-Heinemann for IADMFR 0250-832 x 91/010017-08
Dentomaxillofac. Radiol., 1991, Vol. 20, February
17
Digital image processing in microsialography: K. Yoshiura et al. microsialograms were digitized by a drum-scanning microdensitometer as 512 x 512 x 8-bit images. Both sampling pitch and aperture size were 50,um. Images were transferred to a microcomputer (MicroVax-Il. DEC, USA) for analysis. The grey levels of all the digitized images were corrected by means of reference stepwedges and the densities equalized using the second polynomial equation by least squares approximation. In the digitized images, grey levels 0 and 255 represented complete black and complete white respectively; the intermediate densities were divided into 254 equal levels. Four cut-off points were selected by comparing all the microsialograms of the control group subjectively and set at 16, 32, 80, 128. Each image was partitioned on this basis into five revised grey-level bands: revised grey 0 (original grey 0-15), revised grey 1 (16-31), revised grey 2 (32-79), revised grey 3 (80-
5mm
20mm
..... c
o u
Q)
:0 e o
u
I
IOmm
l/l
3mm
Figure I Diagram showing the radiation field for the right parotid region: all areas of the gland are covered. P. parotid gland
Together with the four overfilled and six irradiated glands, a total of 35 glands were used for the analyses. Between three and six serial microsialograms were obtained of each gland.
a II
Digital image processing Figure 2 summarizes the image analysis procedure. The
s:
10 9 8
g.7 ~
6
c ~ 5
4
3 2
b'
- - - - - - - - - - - - 1- - - I- - - - - - - I- I- - I- - - I 1- - I I-
-
,-
4 4
I
2' Z Z
-
I-
I
-
-
- - -
I-
- - - - - - I- - - - .- - I -
I
I
I 'I I 'i - - ,- - i -
- - 'i T 2 -
l-I-
-
I-
12 10 >u
8
:> 13
6
eQ)
...
Q)
u.
4 2
Five-level quantitizatian
0
C Run length analysis
Figure 2 Flow diagram showing the sequence of image processing
18
Dentomaxillofac. Radiol., 1991, Vol. 20, February
2
3
4
5
6
Run length
Figure 3 Diagram to show run-length analysis. a. Binary figure simulating the duct. b. Run-length matrix. c. Histogram of the run lengths
Digital image processing in microsialography: K. Yoshiura et al. 127) and revised grey 4 (128-255). Revised grey 0 and 4 indicate complete black and white. These five bands defined the background, soft tissue, peripheral minor, middle-sized and major ducts on the microsialograms. Each image was then assessed for the run length of each of these grey levels which was displayed as the run-length histogram. The run length is the number of juxtaposed picture elements having the same grey level in one scan direction 16. For example, the run-length histogram of the binary figure shown in Figure 3a can be obtained by the following procedure: starting with the left-hand column, there are four fragments with a run length of 1. Moving to the next column on the right, there are again four fragments of one run length, and in the next two of two each. The overall matrix of the run length is shown in Figure 3b and the resulting histogram in Figure 3c. Run length was measured by scanning the image vertically in relation to the main excretory duct. The mean value and standard deviation of each value in the histogram was calculated together with the long and short runs emphasis!". Of these four parameters, long runs emphasis demonstrated slight differences in the duct pattern most clearly and was therefore selected as representative of the run-length histogram. It was calculated using the following equation:
where I = run-length value
F(/)
= frequency
of each run length.
The long runs emphasis was calculated for each revised grey level and also compared with the greyscale histogram. Background (grey level 0) was not analysed. Student's t test was used for the statistical evaluation.
Results Relationship between microsialographic appearance, amount of the contrast medium infused, grey-scale histogram and run-length analysis Typical control group microsialograms and 5-grey-level images are shown in Figure 4 and the related grey-scale histogram and run-length analysis in Figure 5. The grey-scale histogram shows that there are significant changes at grey level 3, and, to a larger extent, levels 1 and 4, following the injection of 0.5 and 1.0,u1. The run-length analyses, on the other hand, failed to demonstrate a similar increase in long runs emphasis at grey levels 3 and 4. At the latter, the only significant difference was on injection of 0.5,u1 and the former, with both 0.5 and 1.0,u1. In contrast, significant changes were observed at grey level 2.
Figure 4 Typical microsialograms (above) and their 5-level images (below) of a control parotid gland. The volumes of contrast medium injected are, from left to right, 0.5, 1.0, 1.5 and 2.0pl
Dentomaxillofac. Radiol., 1991, Vol. 20, February
19
Digital image processing in microsialography: K. Yoshiura et al. 15000
500
400 10000
** * •
..f/-\-\
(/)
'iii
(/)
0
s: 300 0-
Q)
>
