Mammogram Analysis – Tumor classificationBackgroundMicrocalcificationsMethods ..Method I - SVMMethod IIResultsResults - ROCMammogram Analysis Mammogram Analysis – Tumor classification– Tumor classification- Geethapriya - Geethapriya RaghavanRaghavanBackgroundBackgroundMammogram – Mammogram – X-Ray image (of gray levels) of inner X-Ray image (of gray levels) of inner breast tissue to detect cancerbreast tissue to detect cancerShows the levels of contrast characterizing Shows the levels of contrast characterizing normal tissue and vesselsnormal tissue and vesselsIssues –Issues –Detect abnormalities (tumors)Detect abnormalities (tumors)Diagnosis - Classify as benign or malignantDiagnosis - Classify as benign or malignantRemove noiseRemove noiseMicrocalcificationsMicrocalcificationsMammograms obtained from MIAS databaseMethods ..Methods ..Non-linear classifiers preferred over linear classifiers Non-linear classifiers preferred over linear classifiers given the randomness in occurrence of tumor cellsgiven the randomness in occurrence of tumor cellsContemporary methods - supervised learning Contemporary methods - supervised learning problem (Wei problem (Wei et al., et al., 2005)2005)Support Vector Machines (SVM)Support Vector Machines (SVM) (Vapnik (Vapnik et al., et al., 1997)1997)Kernel Fisher Discriminant (KFD)Kernel Fisher Discriminant (KFD)Relevance Vector Machines (RVM)Relevance Vector Machines (RVM)Method I - SVMMethod I - SVMSVM was used by Chang SVM was used by Chang et al.,et al., on US images on US images Texture feature – Texture feature – microcalcification area, contrastmicrocalcification area, contrast..Software – SVM Light (Software – SVM Light ((http://svmlight.joachims.org/)(http://svmlight.joachims.org/)The best fitting hyperplane f(x) = wThe best fitting hyperplane f(x) = wT . T . x + b forms the x + b forms the boundaryboundaryFor non-linear SVM, the ‘x’ in the above equation is For non-linear SVM, the ‘x’ in the above equation is replaced by a nonlinear function of ‘x’. replaced by a nonlinear function of ‘x’.Method IIMethod IIUse of wavelet transform to decorrelate data Use of wavelet transform to decorrelate data (image) (Borges (image) (Borges et al.,et al., 2001) 2001)Obtain wavelet coefficients as featuresObtain wavelet coefficients as featuresNormalize coefficients and feed into Nearest Normalize coefficients and feed into Nearest Neighborhood classifierNeighborhood classifierWavelet decomposition - Low frequency Wavelet decomposition - Low frequency coefficients extracted at coefficients extracted at twotwo levels and NNR run levels and NNR run with with euclidean distanceeuclidean distance as metric. as metric.ResultsResults Classifier Microcalcification Contrast Microcalcification AreaNon-linear SVM 67.7 % 78 %Linear SVM 42.8 % 70.4 %NNR 72 % 76.2 %Results - Results - ROCROC Sensitivity = Number of True Positive Sensitivity = Number of True Positive ClassificationsClassifications Number of Malignant LesionsNumber of Malignant LesionsSpecificity = Number of True Negative Specificity = Number of True Negative ClassificationsClassifications Number of Benign LesionsNumber of Benign LesionsSensitivity (y) vs. Specificity (x)Sensitivity (y) vs. Specificity (x)Dotted = lower boundDotted = lower boundRed line = Wavelets + NNRRed line = Wavelets + NNRBlack curve = linear SVMBlack curve = linear SVM specificity0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
View Full Document