- Email: [email protected]
Computerized texture analysis of mammographic parenchymal patterns of digitized mammograms
International Congress Series 1268 (2004) 878 – 881
www.ics-elsevier.com
Computerized texture analysis of mammographic parenchymal patterns of digitized mammograms Hui Li a, Maryellen L. Giger a,*, Olufunmilayo I. Olopade b, Anna Margolis a, Li Lan a, Ioana Bonta a a
Department of Radiology, The University of Chicago, 5841 S. Maryland Ave., MC2026, Chicago, Illinois 60637, USA b Department of Hematology and Oncology, The University of Chicago, Chicago, Illinois 60637, USA
Abstract. This study involved a comparative evaluation of various computer-extracted texture features of mammographic parenchymal patterns of women with BRCA1/BRCA2 gene mutations and those of women at low risk of developing breast cancer. With our method, computerized texture analysis is performed on a region of interest (ROI) within the mammographic image. We analyzed mammograms from 172 subjects: 30 women with the BRCA1 or BRCA2 gene mutation and 142 low-risk women. A series of texture features were automatically extracted from each ROI to assess the mammographic parenchymal patterns in the images. Receiver Operating Characteristic (ROC) analysis was used to assess the performance of the computerized texture features in the task of distinguishing between gene mutation carriers and low-risk subjects. Results show that gene mutation carriers and low-risk women have different mammographic patterns. For the gene carriers, their mammographic images have coarse texture, which is indicated by high coarseness values, low fractal dimension, and low entropy. It is expected that women identified as high risk by such radiographic markers of risk might potentially be more aggressively screened for breast cancer. D 2004 CARS and Elsevier B.V. All rights reserved. Keywords: Image analysis; Mammographic parenchymal patterns; Texture analysis
1. Introduction Image texture is known to provide the rich visual information to humans and is a key component in image analysis. Texture analysis is an important and useful tool in many areas, especially in medical imaging research [1 –9]. There is universal agreement that the human visual system has difficulty in the discrimination of higher-order statistics texture information and frequency spectral properties of an image. Computerized texture analysis has received growing interest in applications for medical imaging in recent years. It has been used to extract clinical meaningful information from various medical imaging modalities. In this study, various texture features were extracted by the computer from * Corresponding author. Tel.: +1-773-834-5099; fax: +1-773-702-0371. E-mail addresses: [email protected] (H. Li), [email protected] (M.L. Giger). 0531-5131/ D 2004 CARS and Elsevier B.V. All rights reserved. doi:10.1016/j.ics.2004.03.212
H. Li et al. / International Congress Series 1268 (2004) 878–881
879
digitized mammograms. Then, a comparative evaluation of the mammographic texture features was conducted using the parenchymal patterns of women with BRCA1/BRCA2 gene mutations and women at low risk of developing breast cancer. 2. Materials and methods In this study, we analyzed mammograms from 172 subjects: 30 women with the BRCA1 or BRCA2 gene mutation and 142 low-risk women. Because age is a very important risk factor, the study was also performed on 30 BRCA1/BRCA2 gene mutation carriers and 60 low-risk women who were randomly selected from the 142 low-risk subjects and were age-matched to the 30 gene mutation carriers at 5-year intervals [10]. Mammograms were digitized using a Konica laser scanner at 0.1 mm pixel size and 10-bit gray-level quantization. Regions of interest (ROIs), 256 pixels by 256 pixels in size, were manually selected from the central breast region immediately behind the nipple [11] (see Fig. 1). The ROIs were used in subsequent computerized feature extraction to assess mammographic parenchymal patterns. The texture features were automatically extracted from each ROI to assess the mammographic parenchymal patterns in the images. These texture features can be grouped into six categories: (1) features based on absolute values of gray levels, (2) features based on gray-level histogram analysis, such as balance and skewness, (3) features based on spatial relationships among gray-levels, such as contrast, coarseness, and features extracted from co-occurrence matrices, (4) features based on fractal analysis, (5) features based on edge frequency, and (6) features based on Fourier transform analysis. Receiver Operating Characteristic (ROC) analysis [12] was used to assess the performance of the computerized texture features in the task of distinguishing between gene mutation carriers and low-risk subjects. 3. Results Quantitative texture analysis on normal digitized screen/film mammograms demonstrated that gene mutation carriers and low-risk women have different mammographic
Fig. 1. A sample ROI selected from the central breast region behind the nipple in a digitized mammogram.
880
H. Li et al. / International Congress Series 1268 (2004) 878–881
Table 1 Performance of individual features in distinguishing between BRCA1/BRCA2 gene mutation carriers and the low-risk cases in the entire group and the age-matched group (Az is area under ROC curve) Features
Entire group Az F S.D.
Age-matched group Az F S.D.
Skewness Balance Coarseness Contrasta Energy Entropy Contrastb DBC(f)c DM(f)d Mean gradient FMPe RMSf
0.72 F 0.05 0.68 F 0.05 0.79 F 0.04 0.72 F 0.05 0.66 F 0.05 0.66 F 0.05 0.86 F 0.03 0.74 F 0.04 0.84 F 0.04 0.78 F 0.04 0.75 F 0.04 0.74 F 0.04
0.74 F 0.05 0.70 F 0.06 0.80 F 0.05 0.74 F 0.05 0.68 F 0.06 0.67 F 0.06 0.86 F 0.04 0.77 F 0.05 0.84 F 0.05 0.80 F 0.05 0.74 F 0.04 0.68 F 0.04
a
Calculated using neighborhood gray-tone difference matrix. Calculated using co-occurrence matrices. c Fractal dimension extracted with box-counting algorithm. d Fractal dimension extracted with Minkowski technique. e First moment of power spectrum. f Root mean square of power spectrum. b
parenchymal patterns. The performance from ROC analysis of the computer-extracted texture features in distinguishing between the two groups is listed in Table 1. The skewness measure, which is related to the mammographic density in the breast, yielded Az values of 0.72 and 0.74 in distinguishing between the gene mutation carriers and the low-risk women from ROC analysis in the entire database and the age-matched group, respectively. An Az value of 0.84 is obtained with the Minkowski fractal analysis [DM(f)], for using either the entire database or the age-matched group. The contrast measure calculated from co-occurrence matrices, which is used to evaluate local image variation, yielded an Az value of 0.86 in distinguishing between two groups using either dataset. 4. Conclusion Computerized texture analysis of mammograms can provide radiographic measures of the mammographic parenchymal patterns. These computer-extracted features may be useful for identifying women at high risk for breast cancer and for monitoring the treatment of breast cancer patients. Acknowledgements We would like to thank Zhimin Huo, PhD, for initial studies on the database and Dulcy E. Wolverton, MD, for reviewing the mammograms. This work was supported in parts by a grant from the U.S. Army medical research and Material Command grant (DAMD 98-1209). M.L. Giger and L. Lan are shareholders in R2 Technology (Sunnyvale, CA). It is the policy of the University of Chicago that investigators
H. Li et al. / International Congress Series 1268 (2004) 878–881
881
disclose publicly actual or potential significant financial interests that may appear to be affected by the research activities. References [1] G.D. Tourassi, Journey toward computer-aided diagnosis: role of image texture analysis, Radiology 213 (1999) 317 – 320. [2] J.W. Byng, et al., Automated analysis of mammographic densities, Phys. Med. Biol. 41 (1996) 909 – 923. [3] Z. Huo, et al., Computerized analysis of mammographic parenchymal patterns for breast cancer risk assessment: feature selection, Med. Phys. 27 (2000) 4 – 12. [4] M.R. Chinander, et al., Computerized radiographic texture measures for characterizing bone strength: a simulated clinical setup using femoral neck specimens, Med. Phys. 26 (1999) 2295 – 2300. [5] S. Katsuragawa, K. Doi, H. MacMahon, Image feature analysis and computer-aided diagnosis in digital radiography: detection and characterization of interstitial disease in digital chest radiographs, Med. Phys. 15 (1988) 311 – 319. [6] G.D. Tourassi, et al., Multifractal texture analysis of perfusion lung scans as a potential diagnostic tool for acute pulmonary embolism, Comput. Biol. Med. 31 (2001) 15 – 25. [7] J.M. Mathias, P.S. Tofts, N.A. Losseff, Texture analysis of spinal cord pathology in multiple sclerosis, Magn. Reson. Imaging 42 (1999) 929 – 935. [8] D.J. Eng, et al., Texture analysis of human liver, J. Magn. Reson. Imaging 15 (2002) 68 – 74. [9] S. Herlidou, et al., Comparison of automated and visual texture analysis in MRI: characterization of normal and diseased skeletal muscle, Magn. Reson. Imaging 17 (1999) 1393 – 1397. [10] Z. Huo, et al., Computerized analysis of digitized mammograms of BRCA1 and BRCA2 gene mutation carriers, Radiology 225 (2002) 519 – 526. [11] H. Li, et al., Computerized analysis of mammographic parenchymal patterns for assessing breast cancer risk: effect of ROI size and location, Med. Phys. 31 (2004) 549 – 555. [12] C.E. Metz, ROC methodology in radiologic imaging, Invest. Radiol. 21 (1986) 720 – 733.
www.ics-elsevier.com
Computerized texture analysis of mammographic parenchymal patterns of digitized mammograms Hui Li a, Maryellen L. Giger a,*, Olufunmilayo I. Olopade b, Anna Margolis a, Li Lan a, Ioana Bonta a a
Department of Radiology, The University of Chicago, 5841 S. Maryland Ave., MC2026, Chicago, Illinois 60637, USA b Department of Hematology and Oncology, The University of Chicago, Chicago, Illinois 60637, USA
Abstract. This study involved a comparative evaluation of various computer-extracted texture features of mammographic parenchymal patterns of women with BRCA1/BRCA2 gene mutations and those of women at low risk of developing breast cancer. With our method, computerized texture analysis is performed on a region of interest (ROI) within the mammographic image. We analyzed mammograms from 172 subjects: 30 women with the BRCA1 or BRCA2 gene mutation and 142 low-risk women. A series of texture features were automatically extracted from each ROI to assess the mammographic parenchymal patterns in the images. Receiver Operating Characteristic (ROC) analysis was used to assess the performance of the computerized texture features in the task of distinguishing between gene mutation carriers and low-risk subjects. Results show that gene mutation carriers and low-risk women have different mammographic patterns. For the gene carriers, their mammographic images have coarse texture, which is indicated by high coarseness values, low fractal dimension, and low entropy. It is expected that women identified as high risk by such radiographic markers of risk might potentially be more aggressively screened for breast cancer. D 2004 CARS and Elsevier B.V. All rights reserved. Keywords: Image analysis; Mammographic parenchymal patterns; Texture analysis
1. Introduction Image texture is known to provide the rich visual information to humans and is a key component in image analysis. Texture analysis is an important and useful tool in many areas, especially in medical imaging research [1 –9]. There is universal agreement that the human visual system has difficulty in the discrimination of higher-order statistics texture information and frequency spectral properties of an image. Computerized texture analysis has received growing interest in applications for medical imaging in recent years. It has been used to extract clinical meaningful information from various medical imaging modalities. In this study, various texture features were extracted by the computer from * Corresponding author. Tel.: +1-773-834-5099; fax: +1-773-702-0371. E-mail addresses: [email protected] (H. Li), [email protected] (M.L. Giger). 0531-5131/ D 2004 CARS and Elsevier B.V. All rights reserved. doi:10.1016/j.ics.2004.03.212
H. Li et al. / International Congress Series 1268 (2004) 878–881
879
digitized mammograms. Then, a comparative evaluation of the mammographic texture features was conducted using the parenchymal patterns of women with BRCA1/BRCA2 gene mutations and women at low risk of developing breast cancer. 2. Materials and methods In this study, we analyzed mammograms from 172 subjects: 30 women with the BRCA1 or BRCA2 gene mutation and 142 low-risk women. Because age is a very important risk factor, the study was also performed on 30 BRCA1/BRCA2 gene mutation carriers and 60 low-risk women who were randomly selected from the 142 low-risk subjects and were age-matched to the 30 gene mutation carriers at 5-year intervals [10]. Mammograms were digitized using a Konica laser scanner at 0.1 mm pixel size and 10-bit gray-level quantization. Regions of interest (ROIs), 256 pixels by 256 pixels in size, were manually selected from the central breast region immediately behind the nipple [11] (see Fig. 1). The ROIs were used in subsequent computerized feature extraction to assess mammographic parenchymal patterns. The texture features were automatically extracted from each ROI to assess the mammographic parenchymal patterns in the images. These texture features can be grouped into six categories: (1) features based on absolute values of gray levels, (2) features based on gray-level histogram analysis, such as balance and skewness, (3) features based on spatial relationships among gray-levels, such as contrast, coarseness, and features extracted from co-occurrence matrices, (4) features based on fractal analysis, (5) features based on edge frequency, and (6) features based on Fourier transform analysis. Receiver Operating Characteristic (ROC) analysis [12] was used to assess the performance of the computerized texture features in the task of distinguishing between gene mutation carriers and low-risk subjects. 3. Results Quantitative texture analysis on normal digitized screen/film mammograms demonstrated that gene mutation carriers and low-risk women have different mammographic
Fig. 1. A sample ROI selected from the central breast region behind the nipple in a digitized mammogram.
880
H. Li et al. / International Congress Series 1268 (2004) 878–881
Table 1 Performance of individual features in distinguishing between BRCA1/BRCA2 gene mutation carriers and the low-risk cases in the entire group and the age-matched group (Az is area under ROC curve) Features
Entire group Az F S.D.
Age-matched group Az F S.D.
Skewness Balance Coarseness Contrasta Energy Entropy Contrastb DBC(f)c DM(f)d Mean gradient FMPe RMSf
0.72 F 0.05 0.68 F 0.05 0.79 F 0.04 0.72 F 0.05 0.66 F 0.05 0.66 F 0.05 0.86 F 0.03 0.74 F 0.04 0.84 F 0.04 0.78 F 0.04 0.75 F 0.04 0.74 F 0.04
0.74 F 0.05 0.70 F 0.06 0.80 F 0.05 0.74 F 0.05 0.68 F 0.06 0.67 F 0.06 0.86 F 0.04 0.77 F 0.05 0.84 F 0.05 0.80 F 0.05 0.74 F 0.04 0.68 F 0.04
a
Calculated using neighborhood gray-tone difference matrix. Calculated using co-occurrence matrices. c Fractal dimension extracted with box-counting algorithm. d Fractal dimension extracted with Minkowski technique. e First moment of power spectrum. f Root mean square of power spectrum. b
parenchymal patterns. The performance from ROC analysis of the computer-extracted texture features in distinguishing between the two groups is listed in Table 1. The skewness measure, which is related to the mammographic density in the breast, yielded Az values of 0.72 and 0.74 in distinguishing between the gene mutation carriers and the low-risk women from ROC analysis in the entire database and the age-matched group, respectively. An Az value of 0.84 is obtained with the Minkowski fractal analysis [DM(f)], for using either the entire database or the age-matched group. The contrast measure calculated from co-occurrence matrices, which is used to evaluate local image variation, yielded an Az value of 0.86 in distinguishing between two groups using either dataset. 4. Conclusion Computerized texture analysis of mammograms can provide radiographic measures of the mammographic parenchymal patterns. These computer-extracted features may be useful for identifying women at high risk for breast cancer and for monitoring the treatment of breast cancer patients. Acknowledgements We would like to thank Zhimin Huo, PhD, for initial studies on the database and Dulcy E. Wolverton, MD, for reviewing the mammograms. This work was supported in parts by a grant from the U.S. Army medical research and Material Command grant (DAMD 98-1209). M.L. Giger and L. Lan are shareholders in R2 Technology (Sunnyvale, CA). It is the policy of the University of Chicago that investigators
H. Li et al. / International Congress Series 1268 (2004) 878–881
881
disclose publicly actual or potential significant financial interests that may appear to be affected by the research activities. References [1] G.D. Tourassi, Journey toward computer-aided diagnosis: role of image texture analysis, Radiology 213 (1999) 317 – 320. [2] J.W. Byng, et al., Automated analysis of mammographic densities, Phys. Med. Biol. 41 (1996) 909 – 923. [3] Z. Huo, et al., Computerized analysis of mammographic parenchymal patterns for breast cancer risk assessment: feature selection, Med. Phys. 27 (2000) 4 – 12. [4] M.R. Chinander, et al., Computerized radiographic texture measures for characterizing bone strength: a simulated clinical setup using femoral neck specimens, Med. Phys. 26 (1999) 2295 – 2300. [5] S. Katsuragawa, K. Doi, H. MacMahon, Image feature analysis and computer-aided diagnosis in digital radiography: detection and characterization of interstitial disease in digital chest radiographs, Med. Phys. 15 (1988) 311 – 319. [6] G.D. Tourassi, et al., Multifractal texture analysis of perfusion lung scans as a potential diagnostic tool for acute pulmonary embolism, Comput. Biol. Med. 31 (2001) 15 – 25. [7] J.M. Mathias, P.S. Tofts, N.A. Losseff, Texture analysis of spinal cord pathology in multiple sclerosis, Magn. Reson. Imaging 42 (1999) 929 – 935. [8] D.J. Eng, et al., Texture analysis of human liver, J. Magn. Reson. Imaging 15 (2002) 68 – 74. [9] S. Herlidou, et al., Comparison of automated and visual texture analysis in MRI: characterization of normal and diseased skeletal muscle, Magn. Reson. Imaging 17 (1999) 1393 – 1397. [10] Z. Huo, et al., Computerized analysis of digitized mammograms of BRCA1 and BRCA2 gene mutation carriers, Radiology 225 (2002) 519 – 526. [11] H. Li, et al., Computerized analysis of mammographic parenchymal patterns for assessing breast cancer risk: effect of ROI size and location, Med. Phys. 31 (2004) 549 – 555. [12] C.E. Metz, ROC methodology in radiologic imaging, Invest. Radiol. 21 (1986) 720 – 733.