Patent Grant
9305204
The present invention relates to a method and apparatus to detect the microcalcifications (MC) in scanned images including X-ray images, computer tomography etc. More particularly, the present invention is directed to a method for detecting MC in X-Ray images such as mamograms with varied densities using nonlinear energy operator (NEO) to favor precise detection of early breast cancer. Microcalcifications, in mammogram are associated with both high intensity and high frequency content. The same NEO output is useful to detect the irrelevant curvilinear structures(CLS) thereby helps in reducing the false alarms in microcalcification detection technique. The proposed methodology is found to be effective on different dataset (scanned film, mammograms with large spatial resolution such as CR and DR) of varied breast composition (viz. dense, fatty glandular, fatty). It is further demonstrated quantitatively involving Free-response receiver operating characteristic
(FROC). Importantly the method and apparatus of the invention can be used in conjunction with machine learning techniques viz. SVM to favour incipient or small microcalcifications and enabling early detection of breast cancer and reducing death rates. Advantageously the method and apparatus are found to benefit further radiologists for confirming detection of micro-calcifications in X-rays and the like.
Microcalcifications(MCs) are tiny spots of calcium deposits that can be scattered or clustered throughout the tissue as found in 2D and 3D images. Singled out MCs indicate the presence of tiny benign cysts, but their clustered appearance signify the presence of early cancer. When found on a mammogram, MCs simply indicate the presence of early breast cancer. The early delectation of MC will aid in reducing the rate of mortality in women. Hence, an adequate and robust MC detection technique is necessary to increase the performance of breast cancer diagnosis. However, identification of MC had been a challenging task to the researchers. This is due to a large variety of breast composition, highly textured breast anatomy and inherent low contrast of mammogram in spite of the recent advances of technology. Moreover MC is very small in size. In mammograms, they appear as a tiny bright spots. Along with MC, different types of breast structures (like curvilinear structures) may appear. It increases the chance of false positives in MC detection technique.
In recent periods, researchers have tried different image processing, machine learning as well as statistical approaches to single out MCs in mammogram image. Nishikawa et al. (1995) introduced a difference technique to enhance MC followed by morphological erosion to reduce false positives. In difference technique approach two consecutive filters are introduced; one for enhancing the MCs and the later for suppressing them [Nishikawa, R. M., Giger, M. L., K. D. C. J. V., and Schmidt, R. A., “Computer-aided detection of clustered micro-calcifications on digital mammograms”, Medical and Biological Engineering and Computing, Springer Berlin/Heidelberg, 1994, pp. 174-178]. Chang et al. (1998) proposed a fuzzy logic based MC detection approach followed by curvilinear structure removal using curve detector to reduce false positive percentages [Cheng, H, Y. M. L., and Freimanis, R. I., “A Novel Approach to Micro-calcification Detection Using Fuzzy Logic Technique,” IEEE Transactions on Medical Imaging vol. 17, issue 3, 1998, pp. 442-450]. Gurcan et al. (1999) presented nonlinear sub band decomposition technique for MC detection [Gurcan, M. N., Y. Y. and cetin, A. E., “Micro-calcification segmentation and mammogram image enhancement using nonlinear filtering”, Proceedings of the IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing (NSIP'99), Antalya, Turkey, 1999, June 20-23]. Melloul et al. (2002) developed an entropy thresholding based approach for MC detection. In this two stage technique, multi-scale morphological operation is initially introduced for removing background tissue followed by entropy thresholding based on 3D co-occurrence matrix for locating MCs [Melloul, M. and Joskowicz, L. “Segmentation of micro-calcification in X-ray mammograms using entropy thresholding”, Technical report, In Proceedings of the 16th International Congress on Computer-Assisted Radiology and Surgery, 2002]. Naga et al. (2002) demonstrated that classification efficiency could be farther improved by applying successive enhancement learning based support vector machine approach (SVM-SEL) [Naga, I., Yang, Y., Wernick, M. N., Galatsanos, N. P., and Nishikawa, R. M., “A Support Vector Machine Approach for Detection of Microcalcifications”, IEEE Trans. on Med. Imaging, vol. 21, December 2002, pp. 1552- 1563]. Guan et al. (2008) introduced scale invariant feature transform (SIFT) to locate MCs [Guan, Q., Zhang, J., S. C. and Todd-Pokropeka, A., “Automatic Segmentation of Micro-calcification Based on SIFT in Mammograms”, BMEI ‘08: Proceedings of the 2008 International Conference on Biomedical Engineering and Informatics’, IEEE Computer Society, Washington, DC, USA, 2008, pp. 13-17]. Normalized Tsallis entropy and Type II fuzzy set technique has been introduced by Mohanalin et al (2009) in MC detection [Mohanalin, J., P. K. K. and Kumar, N., “Micro-calcification Segmentation Using Normalized Tsallis Entropy: An Automatic “q” Calculation by Exploiting Type II Fuzzy Sets,” IETE Journal of Research, vol. 55, 2009, pp. 90-96]. According to literature, the machine learning based MC detection method seems to achieve the best accuracy [Jinshan, T., Rangayyan, R. M., Jun, X., El Naga, I., and Yongyi, Y.; “Computer-Aided Detection and Diagnosis of Breast Cancer with Mammography: Recent Advances”, IEEE Trans. on Information Technology in Biomedicine, 2009, pp. 236-251].
Along with MC detection, different types of breast structures (like curvilinear structures) may appear as false positives. In literature, curve detector [2], phase congruency [Linguraru, M. G., Marias, K., Ruth, E., and Brady, M., “A biologically inspired algorithm for micro-calcification cluster detection”, Medical Image Analysis Elsevier, vol. 10, Issue 6, pp. 850-862, December 2006] etc. has been tried out for curvilinear structure removal. It enhances the quality of MC detection algorithm.
There has been thus a need in the art to developing a method and apparatus for MC detection in 2D and 3D with a background of large variety of breast composition with highly textured breast anatomy and inherent low contrast of mammogram, ensuring very high true positive(TP) and low false positive(FP) simultaneously in the detection process. In the present work, a novel and automated MC detection technique based on Nonlinear Energy
Operator (NEO) is presented after breast region segmentation. The same NEO output will help in curvilinear structure removal.
It is thus the basic object of the present invention to provide for a method and apparatus for low cost, precise and more reliable detection of microcalcification (MC) in X-ray images including 2D and 3D using non linear energy operator(NEO).
A further object of the present invention is directed to a method and apparatus for detection of microcalcification (MC) as symptom of early stage of breast cancer for variety of breast composition including highly textured breast anatomy overcoming the limitations of low contrast of mammogram especially with varied densities.
A still further object of the present invention is directed to a method and apparatus for detection of microcalcification (MC) adapted to reduce chance of miss and false positives (FP) and ensure very high true positive (TP) in MC detection.
A still further object of the present invention is directed to a method and apparatus for detection of microcalcification (MC) wherein said method is capable of highlighting incipient MC.
A still further object of the present invention is directed to a method and apparatus for detection of microcalcification (MC) wherein the same NEO output of the image can help in curvilinear structure(CLS) removal for further refining of the results.
A still further object of the present invention is directed to a method and apparatus for detection of microcalcification (MC) wherein NEO in conjunction with SVM based learning method can be applied to further reduce the FP while maintaining high TP fraction.
A still further object of the present invention is directed to a method and apparatus for detection of microcalcification (MC) wherein a MC identification algorithm is implemented in real time software which can reduce FP fraction without sacrificing high TP fraction.
The basic aspect of the present invention is thus directed to a method to detect the microcalcifications(MC) in images/mammograms comprising the steps of
A further aspect of the present invention is directed to said method wherein non-linear energy operator(NEO) based microcalcification segmentation step comprises (i) pre emphasize the calcification spikes in row as well as in column directions in 2D image and/or in three axis directions for 3D image and generate said smoothed Nonlinear Energy Operator (SNEO); and (ii) determine the spike positions by applying the optimal threshold locally for each row and column separately.
A still further aspect of the present invention is directed to said method, wherein said step of pre emphasizing calcification spikes comprising (i) non linear operations on the signal known as Non Linear Energy Operator(NEO) using simply three neighbouring samples of signal to calculate its energy given by
φ[x(n)]=x(n)2−x(n−1)x(n+1) (1)
where x(n) is a simple one dimensional signal and φ[x(n)] is energy output, and wherein output of NEO is the instantaneous energy of high pass filtered version of signal; and (ii) introducing a smoothing function preferably window function with NEO to improve the negative terms present in energy signal followed by repeated convolution of non linear energy operator with smoothing function to yield said smoothed non linear energy operator(SNEO), given by
φ′[x(n)]=φ[x(n)]w(n) (2)
where w(n) is the window function and applying SNEO independently to rows and to columns of an image, two energy images are obtained for both row and column directions whereby SNEO highlights the high frequency and high amplitude content of signal with the instantaneous nature of SNEO enhancing the MC spikes.
A still further aspect of the present invention is directed to said method, wherein energy based thresholding is applied to individual rows(or columns) of the energy image to obtain the optimal threshold for detecting suspicious regions where the Information gain due to the partitioning of a row or a column A of an energy image having entropy E(A) into two groups, the background (A1) and the foreground (A2) pixels, by any energy intensity level T between the maximum and minimum energy intensity levels of A, and wherein size of group A1 and A2 is adjusted by changing the value T from minimum level to maximum level of A, computing background entropy ET (A1) and foreground entropy ET (A2) for the groups A1 and A2, said gain denoted by
where, Nk, k=1,2 are the pixel count of foreground or background respectively and N is the total pixel of A, and according to information theory, the optimal threshold is obtained as the value of T that yields the maximum Gain(T)
Yet another aspect of the present invention is directed to said method, wherein any row or column(A) of an energy image is binarized with calcification area as foreground, said binarization operation executed independently for all row and columns, placing said binarized individual rows (or columns) in chronological order of row (or column) to form a binarized image, denoted by Brow (or Bcol) and performing Boolean addition between these binarized images to achieve the final image/result.
A still further aspect of the present invention is directed to said method, wherein said step of false positive reduction comprising removal of curvilinear structure texture appearances of image/mammogram based on the nonlinear energy operator comprises introducing a parameter Φ to detect the CLS elements, given by
reducing the average threshold Tavg to p % of average threshold (where p<100), calculating Φ of each element, and if Φ is lower than certain threshold (say Tφ), object is likely to be calcification as its compactness is high and elements having higher Φ are treated as CLS elements and thus discarded.
Importantly, in said method, simultaneous presence of peaks in both the row and column and/or depth for 3D images indicates the presence of MC.
According to a preferred aspect of the present invention directed to said method wherein said images/ /mammograms include scanned film images, X-Ray, mammograms, computer tomography 2D or 3D of different types used to detect the microcalcifications(MC).
A further aspect of the present invention is directed to a method of early detection of breast cancer comprising:
A still further aspect of the present invention is directed to a system to detect the microcalcifications (MC) in images/mammograms following the above said method comprises
A still further aspect of the present invention is directed to said system to detect the microcalcifications(MC) in images/mammograms comprising computer aided diagonistic (CAD) system adapted for
The objects and advantages of the present invention are described with reference to the following accompanying illustrative figures.
The present invention relates to a method and apparatus to detect the microcalcifications (MC) in X-ray images e.g. MC in mammograms with varied densities using nonlinear energy operator (NEO) to favor precise detection of early breast cancer.
Microcalcifications (MC) are tiny spots of calcium deposits that can be scattered or clustered throughout the tissues as found in 2D and 3D images. When found on a mammogram, MCs simply indicate the presence of early breast cancer. Their early detection will aid in reducing mortality rates. However identification of MC is a challenging task to the researchers due to very small size of MC with a background of large variety of breast composition, highly textured breast anatomy and inherent low contrast of mammogram. Along with MC, different types of breast structures (like curvilinear-structures) may appear. It increases the chance of miss and false positives (FP) in MC detection. We have devised a new technique to identify the MCs with very high true positive (TP) and low false positive (FP) simultaneously.
The present invention is directed to developing a new method to detect MC using nonlinear energy operator (NEO) wherein Binarized output of NEO (based on optimal threshold) signifies the presence of MC. If both row and column wise NEO indicates presence of peak, it is a probable MC candidate in 2D images. For 3D images NEO has to be applied in all the three axes. The threshold selection is data driven (using entropy gain). The same NEO output of the image can help in curvilinear structure (CLS) removal. Importantly, using machine learning technique (viz. SVM) in conjunction with NEO based detection, the FP can be further reduced while maintaining high TP fraction. Proposed technique is capable of highlighting incipient MC.
The method of detection of MC involves nonlinear energy operator (NEO) for MC enhancement in digital X-ray image. Probable MC candidates are detected from NEO output of each row and column (and depth for 3D images) using data driven optimal (information gain based) threshold. Simultaneous presence of peaks in both the row and column (and depth for 3D images) indicates the presence of MC. The same NEO output helps to refine the result by CLS removal. Machine learning is used in conjunction with NEO leading to more refinement of the result (with less FP and high TP). Advantageously, The method of MC detection using NEO and machine learning aided NEO technique are more efficient in detecting incipient or small microcalcifications.
Microcalcification Detection
The present invention is directed to developing an algorithm to detect MC in mammograms with varied densities. The algorithm includes three major steps; preprocessing, segmentation and false positive reduction. In preprocessing stage, all non-breast artifacts (like human introduced labels, radiopaque artifacts) are eliminated followed by down sampling the mammograms to reduce computational complexity. In the following step a novel segmentation algorithm, Nonlinear Energy Operator (NEO) is introduced to single out the suspicious regions for further analysis. Finally, NEO is efficiently reused to exclude the irrelevant curvilinear structures to enhance the quality of detection.
Reference is first invited to the accompanying
Preprocessing
Automated analysis of mammograms may be affected by the presence of human introduced labels and radiopaque artifacts present in non-breast regions. Thus, in order to limit the search for abnormal findings within the breast region, mammogram segmentation is a necessary step as preprocessing. The approach adopted to segment the breast region in the present invention is named Shape-based Breast Segmentation (SBS) as shown in accompanying Figure known in the existing art [Seth, S., Mukhopadhyay, S., Das, P., and Bhattyacharyay, P., “Shape-based Breast Segmentation in Mammograms”, IETE Journal of Research, likely Issue Mar-April 2011. (Accepted)]. Moreover to reduce the computational complexity, mammograms with large spatial resolution (viz. CR and DR) are down sampled.
Non-linear Energy Operator Based Microcalcification Segmentation
MCs are very tiny high intensity spots appear in mammograms. It shows spike like behavior in both the directions of mammogram (in row as well as in column direction). Accompanying
Due to its spike like nature, problem of microcalcification segmentation is reduced to detection of peak in two dimensions. In literature, different peak detection algorithms have been tried out. Approximately, they can be partitioned into two steps; First stage is to pre-emphasize the calcification spikes in both the directions (in row as well as in column directions) and in later stage determine the spike positions by applying the suitable threshold locally (for each row and column separately).
Pre-emphasizing calcification spikes
Kaiser et al. (1990) has derived the expression for energy of a signal in light of an algorithm known as Teager's algorithm [Kaiser, J. F., “On a simple algorithm to calculate the energy of a signal,” in Proc. IEEE Int. Conf. Acoustic Speech and Signal Processing, Albuquerque, NM, pp. 381-384, Apr. 1990.]. The algorithm uses simply three neighbouring samples of signal to calculate its energy.
φ[x(n)]=x(n)2−(n−1)x(n+1) (1)
where x(n) is a simple one dimensional signal and φ[x(n)] is energy output. A close look at the algorithm shows that it involves nonlinear operations on the signal shortly known as NEO, Nonlinear Energy Operator. Output of NEO is the instantaneous energy of high pass filtered version of signal. This instantaneous nature of NEO is perfect for detection of transient part of signal [Mukhopadhyay, S., and Ray, G. C, “A New Interpretation of Nonlinear Energy Operator and Its Efficacy in Spike Detection”, IEEE Transactions on Biomedical Engineering, vol. 45, No. 2, February 1998].
A smoothing function (such as Bartlett window function) is introduced with NEO to improve the negative terms present in Teager's energy signal. Repeated convolution of Teager's nonlinear energy operator with smoothing function yields smoothed Nonlinear Energy Operator (SNEO).
φ′[x(n)]=φ[x(n)]x(n) (2)
where w(n) is the window function. Based on best simulation results, the window function of width 5 is selected in this work. By applying φ′[x(n)]=φ[x(n)]w(n)
It can be shown that SNEO is appropriate in highlighting the high frequency and amplitude content of signal as illustrated in accompanying
Optimal Threshold Selection for Detecting Suspicious Regions
Entropy based thresholding method is applied to individual rows (or columns) of the energy image to obtain the optimal threshold. Let any row (or column) of an energy image be A and its entropy be E (A). T is any energy intensity level between the maximum and minimum energy intensity levels of A. This T divides A in two groups; the background (A1) and the foreground (A2) pixels. The size of group A1 and A2 is adjusted by changing the value T from minimum level to maximum level of A. Background entropy ET (A1) and foreground entropy ET (A2) are computed for the groups A1 and A2. The Information gain due to the partitioning of A is denoted by;
where, Nk, k=1,2 are the pixel count of foreground or background respectively and N is the total pixel of A. According to information theory, the optimal threshold is the value of T that yields the maximum Gain(T) [T. Mitchell, Machine Learning, Tata McGraw Hill, 3rd Chap., ISE Editions, pp. 57-59, 1997].
Using this optimal threshold level the A is binarized with calcification probable area as foreground. This binarization operation is executed independently for all row and columns. Obtained binarized individual rows (or columns) are placed in chronological order of row (or column) to form a binarized image, denoted by Brow (or Bcol). Final result is achieved by performing Boolean addition between these binarized images. Still the obtained image contains a few non calcification regions mainly caused by curvilinear structures of mammogram. Accompanying
False Positive reduction
Presence of curvilinear structures (CLS) causes textured appearance of mammogram. It gives rise to False Positive (FP) percentage of any micro-calcification segmentation algorithm. Hence CLS removal is a vital post processing step for False Positive (FP) reduction. Different types of algorithms like curve detector [2], phase congruency [9] etc. have been used for curvilinear structure removal. Implementing these algorithms for curvilinear structure removal will increase over all computation time of the micro-calcification segmentation algorithm. Hence, a novel approach of curvilinear structure removal based on nonlinear energy operator itself is introduced in this paper which increases computational speed of the technique.
After applying NEO based MC segmentation, each segmented elements (or objects) are treated independently. To detect the CLS elements, a parameter Φ is introduced.
Reducing the average threshold Tavg to p % of average threshold (where p<100), Φ of each element is calculated. If Φ is lower than certain threshold (say Tφ), object is likely to be calcification as its compactness is high. Elements having higher Φ are treated as CLS elements and thus discarded. CLS removed images are shown in accompanying
Dataset Preparation & Machine Configuration:
The proposed MC detection technique is developed using a dataset obtained from varied sources. From Mini-MIAS dataset 22 (7 fatty, 8 fatty glandular and 7 dense glandular) with MC present and 15 (6 fatty, 5 fatty glandular and 4 dense glandular) with normal mammograms are added to the simulation dataset (refer
For reduction of processing time, CR and DR images are down sampled in each direction by a factor of 2 and 4 respectively. This will reduce the CR mammogram resolution to 0.194 mm per pixel and DR mammograms resolution to 0.28 mm per pixel. Original MIAS dataset has already been reduced to 0.2 mm per pixel and named Mini-MIAS dataset. Microcalcifications are manually drawn to prepare the ground truth for CR and DR images. These are confirmed by radiologist (refer
Proposed NEO Based Microcalcification Detection Results
Nonlinear energy operator based micro-calcification segmentation technique is reported in this chapter. This technique also helps in removing the irrelevant curvilinear structures thereby reducing the false alarms present in mammogram. Based on best simulation result, the percentage reduction of average threshold (p) and CLS removal parameters (Tφ) are set as 75 and 15 respectively. Using this parameter values, NEO is applied on varied dataset of mammogram for micro-calcification detection.
Proposed technique is applied on DR and CR mammograms of different types (viz. dense, fatty). The results are shown in
Proposed technique is applied on scanned film mammograms of different types (viz. fatty, fatty glandular and dense glandular). The results are shown in
From the simulation results, it has been found that the proposed micro-calcification detection method has performed excellent for DR, CR and scan filmed mammograms irrespective of its viewing angle and density variation.
Quantitative Comparison of Microcalcification Detection Methods
Several techniques have been tried out for efficient microcalcification detection. While increasing the accuracy false alarms are increasing in alarming rate. The proposed NEO based microcalcification detection method shows supremacy over established microcalcification detection method in terms of execution time and accuracy. It can precisely detect single as well as clusters of microcalcification with minimal amount of false alarms. Object cluster detection rate of NEO based method is below 1 false positive per image without any miss for all types of dataset (viz. Mini-MIAS, CR, and DR). This has been demonstrated using free-Response Operator Characteristic (FROC) plot. With this accuracy NEO seems to be the efficient and accurate microcalcification cluster detection technique as illustrated in accompanying
For the case of single object detection, the NEO based microcalcification detection method accurately detects microcalcification with nearly 5 false positive per image without any miss for all types of dataset (viz. Mini-MIAS, CR, and DR). This has been demonstrated using FROC plot as illustrated in accompanying
Using SVM-based machine learning approach, this accuracy rate can bring under 2 false positive per image without any miss. Here, eccentricity, homogeneity and standard deviation is introduced as features to train the SVM classifier for refining the NEO outputs. Quantitative analysis is made among the microcalcification detection techniques applied on various dataset viz. Mini-MIAS, CR, and DR using FROC plot as illustrated in accompanying
The experimental result in this work has been confirmed by radiologist. The proposed technique has been tested on various dataset viz. Mini-MIAS, CR, and DR with varied size of microcalcifications ranging from 0.2 mm to 0.5 mm. The proposed method shows accurate results in case of small MC structures of size in-between 0.2 to 0.3 mm (area of maximum 2 pixels for CR and Scanned film images and area of 1 pixel for DR images). Proposed MC detection technique shows accuracy rate of 2 false positive per image without any miss. Comparison of the proposed and established Naga method [5] is described using FROC plot for the small MC structures as illustrated in accompanying
Additionally, the execution time of the proposed method is less than 3 second per image. Higher processing efficiency of the proposed technique will help the radiologist in making explicit and fast decision.
It is thus possible by way of the present invention to providing a detection method for Microcalcifications which appear as spike like tiny bright spots in X ray images. By applying a robust spike detection algorithm, these spike like microcalcifications can be figured out. Hence, a novel and automatic approach based on nonlinear energy operator (NEO) has been proposed for microcalcification detection which is extremely sensitive to transient part of signal. Mostly all the microcalcification detection techniques exist in literature undergoes an additional false positive reduction step for elimination of the irrelevant curvilinear structures. It has been shown that the same NEO output will help in microcalcification detection along with its refinement. This will results in saving computation time. The proposed method shows supremacy over established methods by accurately detecting the microcalcification clusters within 1 false alarm per image. For the case of single microcalcification, NEO can accurately detect 5 false positives per image without any miss.
Using SVM based learning method; this result can be further refined leading to 2 false positive alarms per image without any miss. In this work, FROC curves indicate that the NEO based microcalcification detection yields the best performance when compared to several existing method. In case of incipient and small microcalcification structures detection, the proposed technique shows supremacy over existing techniques. The method and apparatus of the present invention is thus capable of providing the following advantageous features in MC detection in X-ray images:
| Number | Date | Country | Kind |
|---|---|---|---|
| 1372/KOL/2010 | Dec 2010 | IN | national |
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/IN2011/000824 | 12/5/2011 | WO | 00 | 6/5/2013 |
| Publishing Document | Publishing Date | Country | Kind |
|---|---|---|---|
| WO2012/077130 | 6/14/2012 | WO | A |
| Number | Name | Date | Kind |
|---|---|---|---|
| 4907156 | Doi et al. | Mar 1990 | A |
| 5598481 | Nishikawa | Jan 1997 | A |
| 5999639 | Rogers | Dec 1999 | A |
| 8977019 | Chan | Mar 2015 | B2 |
| 20050152588 | Yoshida | Jul 2005 | A1 |
| 20060147101 | Zhang | Jul 2006 | A1 |
| Entry |
|---|
| International Search Report dated Mar. 26, 2012 for related PCT App. No. PCT/IN2011/000824. |
| Mukhopadhyay S. et al: “A New Interpretation of Nonlinear Energy Operator and Its Efficacy in Spike Detection”, IEEE Transactions on Biomedical Engineering, IEEE Service Center, Piscataway, NJ, USA, vol. 1. 45, No. 2, Feb. 1, 1998; pp. 180-187. |
| Cheng H. D. et al: “Computer-aided detection and classification of microcalcifications in mammograms: a survey;” Pattern Recognition, Elsevier, GB, vol. 1. 36, No. 12; Dec. 1, 2003; pp. 2967-2991. |
| R. M. Nishikawa, M. L. Giger, K. Doi, C. Vyborny, and R. A. Schmidt, “Computer-aided detection of clustered microcalcifications on digital mammograms”, Med. Biological Eng., Computing, vol. 33, pp. 174-178. 1995. |
| Cheng, H, Y. M. L., and Freimanis, R.I., “A Novel Approach to Micro-calcification Detection Using Fuzzy Logic Technique,” IEEE Transactions on Medical Imaging vol. 17, issue 3, pp. 442-450; 1998. |
| Gurcan, M. N.,Yardimci, Yasmen, and Cetin, A. E., “Micro-calcification segmentation and mammogram image enhancement using nonlinear filtering”, Proceedings of the IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing (NSIP'99) 1999. |
| Melloul, M. and Joskowicz, L. “Segmentation of micro-calcification in X-ray mammograms using entropy thresholding”, Technical report, in Proceedings of the 16th International Congress on Computer-Assisted Radiology and Surgery, 2002. |
| Naqa, I., Yang, Y., Wernick, M. N., Galatsanos, N. P., and Nishikawa, R. M., “A Support Vector Machine Approach for Detection of Microcalcifications”, IEEE Trans. on Med. Imaging, vol. 21, pp. 1552-1563; Dec. 2002. |
| Guan, Q., Zhang, J., s. C. and Todd-Pokropeka, A., “Automatic Segmentation of Micro-calcification Based on SIFT in Mammograms”, BMEI '08: Proceedings of the 2008 International Conference on Biomedical Engineering andInfornnatics, IEEE Computer Society, Washington, DC, USA, pp. 13-17, 2008. |
| Mohanalin, J., P. K. K. and Kumar, N., “Micro-calcification Segmentation USing Normalized Tsallis Entropy: An Automatic “q” Calculation by Exploiting Type II Fuzzy Sets,” IETE Journal of Research, vol. 55, pp. 90-96, 2009. |
| Jinshan, T., Rangayyan, R.M., Jun, X., El Naqa, 1., and Yongyi, Y.; “ComputerAided Detection and Diagnosis of Breast Cancer with Mammography: Recent Advances”, IEEE Trans. on Information Technology in Biomedicine, pp. 236-251; 2009. |
| Linguraru, M. G., Marias, K., Ruth, E., and Brady, M., “A biologically inspired algorithm for micro-calcification cluster detection”, Medical Image Analysis Elsevier, vol. 10, Issue 6, pp. 850-862, Dec. 2006. |
| Number | Date | Country | |
|---|---|---|---|
| 20130251207 A1 | Sep 2013 | US |
