The paper proposes a method for extracting a breast image from a mammogram. For this, mammogram binarization, scaling and segmentation of a binary scaled mammogram with the subsequent selection of the maximum connected component that corresponds to a breast image have been suggested. The proposed binarization uses uniform quantization that simplifies the selection of the threshold value for different mammograms. The proposed binary mammogram scaling uses a fast wavelet transform and an arithmetic mean filter with threshold processing which accelerates further segmentation. The proposed segmentation of binary scaled mammograms uses density clustering to extract connected components that can more accurately extract the breast image. The proposed method for processing a mammogram based on binarization, scaling, and segmentation can be used in various intelligent medical diagnostic systems.
Currently, the methods of automatic and automated diagnostics of microcalcifications
of mammary glands [
For automatic and automated medical diagnostics, the extracting of a breast image from the mammogram, for which the methods of binarization, scaling and segmentation of images can be used, plays an important role.
For binarization of images, usually use such approaches as:
─ automatic selection of a single-level global threshold (for example, Otsu's method)
[
In this regard, it is relevant to create a method of mammogram binarization, which will eliminate these limitations.
For image segmentation, usually use such approaches as:
─ determination of the boundaries of the regions (as the boundaries of the regions,
pixels with a large intensity gradient, as well as differing in color are
selected) [
Taxonomic approach is the most popular of them.
Traditional methods of taxonomic approach are the following:
1. Methods based on partition (partition-based, partitioning-based) or center
(centerbased) (for example, k-means [
Methods of taxonomic approach can also be based on metaheuristics [
These methods have one or more of the following limitations: ─ have high computational complexity; ─ do not allow to emit noise and random emissions; ─ clusters cannot have different shapes and sizes; ─ require the setting of the number of clusters; ─ require the determination of parameter values.
In this regard, it is relevant to create a method of mammogram segmentation, which will eliminate these limitations.
Pre-scaling, which helps to reduce image size, is one of the ways to speed up segmentation.
For scaling images, usually use such approaches as:
─ the method of the nearest neighbor [
These methods have one or more of the following limitations: ─ have high computational complexity; ─ provide low quality of images; ─ require the determination of parameter values.
In this regard, it is relevant to create a method for mammogram scaling, which will eliminate these limitations.
The purpose of the work is to create a method for extracting a breast image from a mammogram based on binarization, scaling and segmentation. To achieve the goal, the following tasks have been set and solved: 1. To create a technique for mammogram binarization based on quantization and threshold processing. 2. To develop a technique for binary mammogram scaling. 3. To create a technique for binary scaled mammogram segmentation based on density clustering. 4. To develop a technique for determining the maximum connected component of a binary scaled mammogram, which corresponds to a breast image. 5. To create a technique for transforming the initial mammogram based on the maximum connected component of a binary scaled mammogram. 6. To conduct a numerical study. 2
Mammogram binarization based on quantization and threshold processing In the paper a uniform scalar quantizer, which is optimal (the root-mean-square error of quantization is minimal) is used, and the quantization step is constant.
The proposed mammogram binarization includes the following steps: 1. Set the 8-bit mammogram s(n1, n2 ) , n1 1, N1, n2 1, N2 . Set the number of quantization levels of pixel values L that is a multiple of the power of two. Set the minimum and maximum pixel values smin (in the case of the 8-bit image smin 1), and smax (in the case of the 8-bit image smax 256 ), respectively. Set the threshold value T . 2. Calculate the boundaries of the quantized pixel values for the corresponding quantization levels 3. Calculate the quantized pixel values for the corresponding quantization levels di smin (smax smin )i , i 0, L .
L d di1 , i 1, L . ri i
2 4. Quantize the 8-bit mammogram in the form r1, Q(s(n1, n2 )) ... ...
d0 s(n1, n2 ) d1 rL , dL1 s(n1, n2 ) dL
, n1 1, N1, n2 1, N2 . 1, Q(s(n1, n2 )) T b(n1, n2 ) 0, Q(s(n1, n2 )) T , n1 1, N1, n2 1, N2 .
3
To scale the mammogram, the methods based on filtering and two-dimensional fast wavelet transform are offered in the paper. 3.1
Mammogram scaling based on arithmetic mean filter and threshold processing The proposed mammogram scaling based on arithmetic mean filter and threshold processing includes the following steps: 1. Set the binary mammogram b(n1, n2 ) , n1 1, N1, n2 1, N2 . Set the scaling parameter P , which determines the length of the square window as 2P . Set the threshold value T . 2. Set the line number of the binary scaled mammogram n1 1 . 3. Set the column number of the binary scaled mammogram n2 1. 4. Calculate the average pixel value in a window of 2P 2P size 1
P P b(l1, l2 ) , 2 2 l1,l2 l1 (n1 1)2P 1, (n1 1)2P 2P , l2 (n2 1)2P 1, (n2 1)2P 2P . 5. Convert the binary mammogram to b (n1, n2 ) 1, (n1, n2 ) T .
0, (n1, n2 ) T 6. If it is not the end of the current line of the binary scaled mammogram, i.e. n2 N2 / 2P , then increase the column number of the current line of the binary scaled mammogram, i.e. n2 n2 1 . Go to step 4. 7. If it is not the last line of the binary scaled mammogram, i.e. n1 N1 / 2P , then increase the line number of the binary scaled mammogram, i.e. n1 n1 1. Go to step 3.
As a result, the binary scaled mammogram is formed. 3.2
Mammogram scaling based on two-dimensional fast wavelet transform The proposed mammogram scaling based on two-dimensional fast wavelet transform includes the following steps: 1. Set the binary mammogram b(n1, n2 ) , n1 1, N1, n2 1, N2 . Set the scaling parameter P that determines the number of decomposition levels. 2. Set the decomposition level number i 1. 3. For each line x , x 0, N1 / 2i1 1, at the current i-th decomposition level, a convolution of this line with transition functions FIR-HPF and FIR-LPF, respectively, is performed: di (x, m) 2 ci (x, m) 2
N2 /2i11 k0 N2 /2i11 k0 ci1(x, k)g(k 2m) , m 0, N2 / 2i 1, ci1(x, k)h(k 2m) , m 0, N2 / 2i 1, where c0 (x 1, y 1) b(x, y) . 4. For each column y , y 0, N2 / 2i 1, at the current i-th decomposition level, a convolution of this column with transition functions FIR-HPF and FIR-LPF, respectively, is performed: 5. If i P , then i i 1. Go to step 1. 6. Convert the values of approximating coefficients to the range of values {0,1} cmin mx,iyn cP (x, y) , x 0, N1 / 2P 1, y 0, N2 / 2P 1, cmax max cP (x, y) , x 0, N1 / 2P 1, y 0, N2 / 2P 1,
x,y b(x 1, y 1) round cP (x, y) cmin , x 0, N1 / 2P 1, y 0, N2 / 2P 1, cmax cmin where round (x) is the x rounding.
As a result, the binary scaled mammogram is formed. 4
Binary scaled mammogram segmentation based on density clustering The proposed segmentation of binary scaled mammogram includes the following steps: 1. Set the binary scaled mammogram b(n1, n2 ) , n1 1, N1, n2 1, N2 , where N1 N1 / 2P , N2 N2 / 2P . Set the size of the pixel neighborhood D (in the case of Moore neighborhood D 9 ). Initialize the pixel marking matrix g(n1, n2 ) 0 , n1 1, N1, n2 1, N2 . Initialize the counter of the current number of connected components c 0 . 2. Set the mammogram line number n1 1 . 3. Set the mammogram column number n2 1. 4. Calculate the current pixel number i (n1 1)N2 n2 . 5. If the i-th pixel is already marked, i.e. g(n1, n2 ) 0 , then go to step 20.
Ui, {e | b(l1 n1, l2 n2 ) 1} , e (l1 n1 1)N2 l2 n2 ,
l1 {1, 0,1}, l2 {1, 0,1} . 7. If not all neighbors of the i-th pixel fall into its neighborhood, i.e. Ui, D , then mark the i-th pixel as noise or random emission, i.e. g(n1, n2 ) 1 . Go to step 20. 8. Increase the counter of the current number of connected components c c 1. 9. Mark the i-th pixel as the c-th cluster, i.e. g(n1, n2 ) c . 10. Create a multitude S Ui, . 11. Extract from the set S the first element, i.e. v s1 , and remove it from the set S , i.e. S S \ {v} . 12. Calculate the coordinates of the v -th pixel in the mammogram
m2 v mod N2 , m1 [(v m2 ) / N2 ] , where – taking the integer part of the number, mod – division modulo. 13. If the v -th pixel has been marked as noise or accidental release, i.e.
g(m1, m2 ) 1 , then mark it as the c -th cluster, i.e. g(m1, m2 ) c . 14. If the v -th pixel is already marked, i.e. g(m1, m2 ) 0 , then go to step 19. 15. Mark the v -th pixel, i.e. g(m1, m2 ) c . 16. Determine the neighborhood of the v -th pixel
Uv, {e | b(l1 m1, l2 m2 ) 1} , e (l1 m1 1)N2 l2 m2 ,
l1 {1, 0,1}, l2 {1, 0,1} . 17. If not all neighbors of the v -th pixel fall into its neighborhood, i.e. Uv, D , then go to step 19. 18. Combine the set S with the neighborhood of the v -th pixel, i.e. S S 19. If the set S still contains pixels, i.e. S 0 , then go to step 11.
Uv, . 20. If it is not the end of the mammogram current line, i.e. n2 N2 , then increase the column number of the mammogram current line, i.e. n2 n2 1 . Go to step 4. 21. If it is not the last line of the image, i.e. n1 N1 , then increase the mammogram line number, i.e. n2 n2 1 . Go to step 3.
As a result, the pixel marking matrix of the segmented binary scaled mammogram is formed.
Determination of the maximum connected component of the binary scaled mammogram, which corresponds to a breast image The proposed definition of the maximum connected component of the binary scaled mammogram includes the following steps: 1. Define the pixel marking matrix g(n1, n2 ) , n1 1, N1, n2 1, N2 , where N1 N1 / 2P , N2 N2 / 2P . Set the number of the connected components c . Initialize the vector of the counters of the connected components dimensions z(n) 0 , n 1, c . 2. Set the line number of the pixel marking matrix n1 1 . 3. Set the column number of the pixel marking matrix n2 1. 4. If the binary pixel refers to the connected component, i.e. g(n1, n2 ) 0 , then increase the size counter of the connected components, i.e. z(g(n1, n2 )) z(g(n1, n2 )) 1 . 5. If it is not the end of the current line of the pixel marking matrix, i.e. n2 N2 , then increase the column number of the current line of the pixel marking matrix, i.e. n2 n2 1 . Go to step 4. 6. If it is not the last line of the pixel marking matrix, i.e. n1 N1 , then increase the line number of the pixel marking matrix, i.e. n1 n1 1. Go to step 3. 7. Determine the number of the maximum connected component c* arg max z(n) , n 1, c .
n
As a result, the number of the maximum connected component of the binary scaled mammogram, which corresponds to a breast image, is determined. 6
Transformation of the original mammogram based on the maximum connected component of the binary scaled mammogram 1. Set the pixel marking matrix g(n1, n2 ) , n1 1, N1, n2 1, N2 , N1 N1 / 2P , N2 N2 / 2P . Set the number of the maximum connected component c* . Set the 8bit mammogram s(l1,l2 ) , l1 1, N1, l2 1, N2 . Set the scaling parameter P . 2. Set the line number of the pixel marking matrix n1 1 . 3. Set the column number of the pixel marking matrix n2 1. 4. Convert the original mammogram to , l1 (n1 1)2P 1, (n1 1)2P 2P , l2 (n2 1)2P 1, (n2 1)2P 2P . 5. If it is not the end of the current line of the pixel marking matrix, i.e. n2 N2 , then increase the column number of the current line of the pixel marking matrix, i.e. n2 n2 1 . Go to step 4. 6. If it is not the last line of the pixel marking matrix, i.e. n1 N1 , then increase the line number of the pixel marking matrix, i.e. n1 n1 1. Go to step 3.
As a result, the mammogram containing only a breast image is formed. 7
In the paper, the proposed method for extracting a breast image from a mammogram is investigated.
The quantization level of 16 and the threshold value of 0.5 are selected.
Fig. 1,a shows the original 8-bit mdb274 mammogram from the standard miniMIAS mammogram database. Image size is 1024x1024 pixels.
Fig. 1,b shows the resulting 8-bit mdb274 mammogram that does not use scaling P 0 .
a) b)
According to the experiments for mammograms from the standard mini-MIAS database, which are presented in Fig. 2, for scaling using the arithmetic mean filter with threshold processing, use the scaling parameter value P 3 .
Such a value of the scaling parameter P , on the one hand, does not lead to significant changes in the shape of a breast image (this is typical for values 4, 5, 6), which impair the visual perception, and, on the other hand, does not lead to a significant slowdown in segmentation (this is typical for values 1, 2). Fig. 2. 8-bit mdb274 mammogram in the case of scaling using the arithmetic mean filter with threshold processing: a) the resulting with P = 1; b) the resulting with P = 2; c) the resulting with P = 3; d) the resulting with P = 4; e) the resulting with P = 5; f) the resulting with P = 6.
According to the experiments for mammograms from the standard mini-MIAS database, which are shown in Fig. 3, for scaling using the fast wavelet transform by means of Daubechies wavelet of length 8 (denoted as db4), use the scaling parameter value P 2 .
Such a value of the scaling parameter P , on the one hand, does not lead to significant changes in the shape of a breast image (this is typical for values 4, 5, 6), which impair visual perception, and, on the other hand, does not lead to a significant slowdown in segmentation (this is typical for values 1, 2).
According to Fig. 2-3, the fast wavelet transform gives smoother edges than the arithmetic mean filter with threshold processing, but truncates a breast image, and also requires the choice of the wavelet and its length, and has greater computational complexity, while the arithmetic mean filter with threshold processing does not require the setting of additional parameters.
The dependence of the segmentation time of the binary mdb274 mammogram on the scaling parameter P is shown in Fig. 4. The experiments have been conducted on a computer with an Intel Pentium Quad-Core processor with a clock frequency of 2.58 GHz.
T , sec P
According to Fig. 4, the dependence of the segmentation time on the scaling parameter is close to exponential one and shows that, starting from P 3 , the segmentation time changes only slightly. 8
Conclusions 1. To solve the problem of improving the quality of medical diagnostics, the corresponding methods of digital image processing have been investigated. The research data show that today, the methods of binarization, scaling and segmentation are actively used to extract a breast image from a mammogram. 2. The created method for extracting a breast image from a mammogram sequentially performs: ─ binarization based on uniform quantization and threshold processing, which allows to more accurately separate the background and mammogram objects, simplifies the selection of the threshold value for different mammograms; ─ binary mammogram scaling based on the arithmetic mean filter with threshold processing and fast wavelet transform, which accelerates further segmentation. To increase the accuracy and speed of scaling experimentally, the scaling parameter is determined; ─ binary scaled mammogram segmentation based on density clustering for extraction of connected components, which allows to more accurately extract a breast image, not take noise and accidental releases into account, extract connected components of different shapes and sizes, not indicate the number of connected components; ─ selection of the maximum connected component that matches a breast image; ─ transformation of the initial mammogram based on the maximum connected component of the binary scaled mammogram. 3. The proposed method for extracting a breast image from a mammogram based on binarization, scaling and segmentation can be used in various intelligent medical diagnostic systems.