Path. Res. Pract. 188,565-569 (1992)
Texture Analysis of Histological Images of Giant Cell Tumor of Bone H. Kuwahara, M.Shimazaki, I. Morikita, Y. Chanoki and M. Sakurai Department of Pathology, Osaka City University Medical School, Osaka, Japan
SUMMARY To gain an objective evaluation of histological sections of giant cell tumors of bone (GCT) and osteosarcomas, microscopic pictures were taken and their grey-tone image measured, using a flying spot scanner and computer. Various values of eight parameters expressing certain characteristical brightness distribution patterns were computed, and comparatively examined among the three groups of benign GCT, malignant GCT and osteosarcoma. As a result, some parameters could facilitate differentiation between the histological images of these bone tumors. Especially, "angular second moment (ASM) ", "Contrast" and "Coefficient of variation (CO V)" were useful ellen for discrimination between malignant and benign GCT. After factor analysis of the values of these parameters, scores of each factor for a number of histological scene images were plotted on a 2-dimensional factor plane. On this plane, which was considered to be a histological feature plane, cases of benign GCT were separated from those of osteosarcoma. Cases of malignant GCT were distributed between the two groups. These results suggest that this method could be lIaluable for computer ellaluation of histological images of benign GCT and osteosarcomas.
Introduction The histological diagnoses of bone tumors have been exclusively conducted only by well-trained pathologists with rich experience. For this reason, those judgments were influenced by subjective factors, depending on personal experience. For objective evaluation, several computer-analyses have been designed to characterize whole cells in smear preparations for hematology and urothelial tumor examination4 ,8. The authors hitherto analyzed black-and-white microphotographs of tumor tissues with a computer and tried to define parameters which could be objectively assessed 1,3,5,6, 7. The purpose of this study is to describe a method of extraction of histological features from GeT and osteosarcomas, and to select an appropriate set of parameters to characterize these bone tumors. © 1992 by Gustav Fischer Verlag, Stuttgart
Material and Methods The precise methods and principles used in this study were essentially the same as described by Kanno et aP. Materials were tissue specimens surgically resected from 10 cases of GCT and 9 cases of osteosarcoma. Four micrometer thick paraffin sections were cut and stained with hematoxylin and eosin. Light microscopic grey-tone pictures of bone tumors were taken at 33 x magnification and printed in 11 x 7.5 cm size. For photographing, two tissue sections were selected for each case, and one determination site was adopted in each section. Obviously degenerated and necrotic regions were avoided, and bone trabeculae, hemorrhagic areas were excluded. Then, using a flying spot scanner(made by Kowa Company, Ltd., Tokyo), microphotographs were scanned with a light spot (170 micrometer in diameter), resulting in scanning spot areas of 160 x 130 pixels. Determinating the reflected light from each pixel, the grey-tone value was expressed numerically in 32 values. These values were 0344-0338/92/0 18 8-05 65 $3 .50/0
566 . H. Kuwahara et al. input to a computer, Melcom 70 (Mitsubishi Electric Corporation, Tokyo), and the images were computed as follows: First, four sizes of unit area were set up. The smallest size of unit area is called "Pixel 1", which is equal to the spot light size. Unit area made of neighboring 4 pixels was named as "Pixel 2". So, Table 1. Parameters
.. )
P (i, j)" R
( J = pi,
R
LL
=
I
I' (i, j)
=
P (j, i)
P (i, j)
J
Angular Second Moment (ASM): fl = lp(i, j)j2
LL i
2
J
Contrast:
I.~.~ ,
31
" f2 L.. nn=O
3
I~IJ~I
p(, j)
Illi- jl
I
=n
Correlation: L L (ij)p (i.j) - [12'
f3 =
_i-'.i_ _ _ _ __ 32
P
X
(i)
2 0 "
= L
i~ I
p(i, j)
32
py (j)
L p(i, j)
=
i-oJ
4
Sum of Squares: Variance f4 = L L (i - [1) 2p(i, j) i
5
Inverse Difference Moment (DM): f5 =
6
j
Li Li 1 + (11 - J')' p(i, j) 64
= L
+ y (k)
kP x
k~ 2
P
X
32
32
L L p(i, j). k = 2.3.·-·64
+ y(k) =
I~
i
7
I
+i
i~I ~
k
Sum Variance: 64
f7
= L
(k - f6) 2pX
+ y(k)
k~2
8
P(1'1) P(2.1) .
M-
P(1,2) 1'(2,2)
( P(32,1) R(32,2)
P(1,32) ) P(2,32) P(i,j)
P(~2,32)
Each matrix clement P (i,j) is defined as the frequency of grey of level i occurring after j occurs. On the basis of the combination of 4 kinds of area size, 6 kinds of jump number, and 4 directions, a total of 96 kinds of matrices were generated for each picture to be evaluated, and from these matrices, the 8 parameters in the Table 1 were calculated. The meaning of each parameter ma y be explained as follows: 1) Angular second moment (ASM) is a measure of monotonicity in the transition type of grey-tone level of an image. Images which are homogeneous or show regularly alternating patterns, show high values of this parameter. 2) The contrast is a measure of the magnitude of local variations in an image. 3) The correlation is a measure of the local grey-tone dependencies. If the areas with similar grey-tone values are arranged in specific direction, this parameter gives large positive values. 4) The sum of squares (SOS) indicates the degree of variation of the grey-tone. 5) The inverse difference moment (IDM) is nearly equal to the inverse of the contrast value. 6) The sum average (SA) indicates whether the whole image is in a light tone or dark tone. 7) The sum variance (SV) represents the variation of the sum of the grey-tone values of two neighbouring areas. 8) The coefficient of variation (COV) is a measure of the relative variation of the grey-tone values.
Results
Sum Average: f6
"Pixel 4" was made of 16 pixels. The grey-tone value for each unit area was defined as the average of grey-tone values of pixels constituting it. Secondly, we adopted the jump method. In the "Jump 0" procedure, the grey-tone values of neighboring two unit areas were compared. In the "Jump 1" procedure, the value of two unit areas separated from each other by one unit area was compared. The "Jump 2", "Jump 3", "Jump 4" and "Jump 5" procedures were also similarly defined. This comparison was made in the 4 directions: horizontal (0°), upper right (45°), vertical (90 0), and lower right (_135°) direction. Thus, the transition probability of the grey-tone values was expressed as 32 by 32 grey-tone cooccurrcnce matrix M in the form 2,
Coefficient of Variation: f8 = O/~l
"P(i,j) represents the number of times in which two areas separated by jump number n have gray-tone values i and j (n = 0,1,2,3,4,5). -_."
Texture Analysis of Histological Images of Giant Cell Tumor of Bone H. Kuwahara, M.Shimazaki, I. Morikita, Y. Chanoki and M. Sakurai Department of Pathology, Osaka City University Medical School, Osaka, Japan
SUMMARY To gain an objective evaluation of histological sections of giant cell tumors of bone (GCT) and osteosarcomas, microscopic pictures were taken and their grey-tone image measured, using a flying spot scanner and computer. Various values of eight parameters expressing certain characteristical brightness distribution patterns were computed, and comparatively examined among the three groups of benign GCT, malignant GCT and osteosarcoma. As a result, some parameters could facilitate differentiation between the histological images of these bone tumors. Especially, "angular second moment (ASM) ", "Contrast" and "Coefficient of variation (CO V)" were useful ellen for discrimination between malignant and benign GCT. After factor analysis of the values of these parameters, scores of each factor for a number of histological scene images were plotted on a 2-dimensional factor plane. On this plane, which was considered to be a histological feature plane, cases of benign GCT were separated from those of osteosarcoma. Cases of malignant GCT were distributed between the two groups. These results suggest that this method could be lIaluable for computer ellaluation of histological images of benign GCT and osteosarcomas.
Introduction The histological diagnoses of bone tumors have been exclusively conducted only by well-trained pathologists with rich experience. For this reason, those judgments were influenced by subjective factors, depending on personal experience. For objective evaluation, several computer-analyses have been designed to characterize whole cells in smear preparations for hematology and urothelial tumor examination4 ,8. The authors hitherto analyzed black-and-white microphotographs of tumor tissues with a computer and tried to define parameters which could be objectively assessed 1,3,5,6, 7. The purpose of this study is to describe a method of extraction of histological features from GeT and osteosarcomas, and to select an appropriate set of parameters to characterize these bone tumors. © 1992 by Gustav Fischer Verlag, Stuttgart
Material and Methods The precise methods and principles used in this study were essentially the same as described by Kanno et aP. Materials were tissue specimens surgically resected from 10 cases of GCT and 9 cases of osteosarcoma. Four micrometer thick paraffin sections were cut and stained with hematoxylin and eosin. Light microscopic grey-tone pictures of bone tumors were taken at 33 x magnification and printed in 11 x 7.5 cm size. For photographing, two tissue sections were selected for each case, and one determination site was adopted in each section. Obviously degenerated and necrotic regions were avoided, and bone trabeculae, hemorrhagic areas were excluded. Then, using a flying spot scanner(made by Kowa Company, Ltd., Tokyo), microphotographs were scanned with a light spot (170 micrometer in diameter), resulting in scanning spot areas of 160 x 130 pixels. Determinating the reflected light from each pixel, the grey-tone value was expressed numerically in 32 values. These values were 0344-0338/92/0 18 8-05 65 $3 .50/0
566 . H. Kuwahara et al. input to a computer, Melcom 70 (Mitsubishi Electric Corporation, Tokyo), and the images were computed as follows: First, four sizes of unit area were set up. The smallest size of unit area is called "Pixel 1", which is equal to the spot light size. Unit area made of neighboring 4 pixels was named as "Pixel 2". So, Table 1. Parameters
.. )
P (i, j)" R
( J = pi,
R
LL
=
I
I' (i, j)
=
P (j, i)
P (i, j)
J
Angular Second Moment (ASM): fl = lp(i, j)j2
LL i
2
J
Contrast:
I.~.~ ,
31
" f2 L.. nn=O
3
I~IJ~I
p(, j)
Illi- jl
I
=n
Correlation: L L (ij)p (i.j) - [12'
f3 =
_i-'.i_ _ _ _ __ 32
P
X
(i)
2 0 "
= L
i~ I
p(i, j)
32
py (j)
L p(i, j)
=
i-oJ
4
Sum of Squares: Variance f4 = L L (i - [1) 2p(i, j) i
5
Inverse Difference Moment (DM): f5 =
6
j
Li Li 1 + (11 - J')' p(i, j) 64
= L
+ y (k)
kP x
k~ 2
P
X
32
32
L L p(i, j). k = 2.3.·-·64
+ y(k) =
I~
i
7
I
+i
i~I ~
k
Sum Variance: 64
f7
= L
(k - f6) 2pX
+ y(k)
k~2
8
P(1'1) P(2.1) .
M-
P(1,2) 1'(2,2)
( P(32,1) R(32,2)
P(1,32) ) P(2,32) P(i,j)
P(~2,32)
Each matrix clement P (i,j) is defined as the frequency of grey of level i occurring after j occurs. On the basis of the combination of 4 kinds of area size, 6 kinds of jump number, and 4 directions, a total of 96 kinds of matrices were generated for each picture to be evaluated, and from these matrices, the 8 parameters in the Table 1 were calculated. The meaning of each parameter ma y be explained as follows: 1) Angular second moment (ASM) is a measure of monotonicity in the transition type of grey-tone level of an image. Images which are homogeneous or show regularly alternating patterns, show high values of this parameter. 2) The contrast is a measure of the magnitude of local variations in an image. 3) The correlation is a measure of the local grey-tone dependencies. If the areas with similar grey-tone values are arranged in specific direction, this parameter gives large positive values. 4) The sum of squares (SOS) indicates the degree of variation of the grey-tone. 5) The inverse difference moment (IDM) is nearly equal to the inverse of the contrast value. 6) The sum average (SA) indicates whether the whole image is in a light tone or dark tone. 7) The sum variance (SV) represents the variation of the sum of the grey-tone values of two neighbouring areas. 8) The coefficient of variation (COV) is a measure of the relative variation of the grey-tone values.
Results
Sum Average: f6
"Pixel 4" was made of 16 pixels. The grey-tone value for each unit area was defined as the average of grey-tone values of pixels constituting it. Secondly, we adopted the jump method. In the "Jump 0" procedure, the grey-tone values of neighboring two unit areas were compared. In the "Jump 1" procedure, the value of two unit areas separated from each other by one unit area was compared. The "Jump 2", "Jump 3", "Jump 4" and "Jump 5" procedures were also similarly defined. This comparison was made in the 4 directions: horizontal (0°), upper right (45°), vertical (90 0), and lower right (_135°) direction. Thus, the transition probability of the grey-tone values was expressed as 32 by 32 grey-tone cooccurrcnce matrix M in the form 2,
Coefficient of Variation: f8 = O/~l
"P(i,j) represents the number of times in which two areas separated by jump number n have gray-tone values i and j (n = 0,1,2,3,4,5). -_."
