Early Detection and Classification of Breast Tumor From Mammography

Abstract

Breast cancer is common disease in females and mortality rate due to this disease is constantly on the rise. Early detection and diagnosis of breast tumor can reduce fatal cases. Traditionally, mammography is a recommended screening test for early diagnosis and detection of breast tumor. A healthcare professional reads the mammogram and he/she diagnoses any abnormality. However at early stage, tumor is invariably small and sometimes human eye cannot detect small obscured masses which may cause high rate of false assumptions including false-positives and false-negatives. Computer aided diagnosis systems proved to be a potent method for identification and classification of breast cancer at initial stage. CAD systems recognize the suspicious patterns in mammogram that might suggest a malignancy and notifies the radiologist, who can then examine the abnormal area present in digital mammogram more carefully. In this thesis, six techniques are introduced to address the limitations present in existing diagnostic procedures. The first method presented in this thesis is named as removal of pectoral muscle based on the topographic map and shape-shifting silhouette. This approach is based on a discrete differentiation operator which is an edge detector and computes an approximation of the gradient of the image intensity function. For refinement purpose, a convex hull technique is also applied. This approach accommodates a wide variety of the pectoral muscle geometries even in distorted mammogram than earlier techniques. To evaluate the efficiency of the proposed technique, visual inspection by the radiologist as well as the calculation of performance metric, presented good agreement. For computation of performance metric, number of pixels in pectoral muscle region of the input scans is calculated as ground truth and compared with the number of pixels in proposed geometry, showing the promising results with minimum risk of bias in breast profile. In the automated detection of breast tumor, significance of the false positive rate is considered more valuable than the false negative rate. The proposed algorithm shows good results in terms of the mean false positive rate 0.99 and the Hausdorff distance 3.52mm. These outcomes indicate that the proposed technique outperformed in removal of the pectoral muscle domain. Second method presented in this thesis is named as segmentation of breast lesions using colored texture. The proposed technique indicates that a fully automated color features approach can be used for removing pectoral muscle and segmenting the breast lesion region, effectively, for improving its analysis process. Color features and textural based mathematical morphology is used for segmentation of the breast mass. Proposed algorithm im- proved performance of mass segmentation at maintaining the good visual integrity and high accuracy rate of 98.00% on MIAS images and 97.00% on DDSM images segmentation. Third automated method is named as Bi-Model processing for early detection of breast tumor in the CAD system. The proposed technique is used for detecting tumor cells in the breast masses. This follows the minimal false assumptions approach to detect and segment the desired region. The breast mass is analyzed in terms of the breast texture. Moreover, morphology and connectivity of pectoral muscle pixels are used to discriminate the required region of interest from the rest of breast body for developing the proposed CAD system. Furthermore, for segmentation purpose a subsequent design based on pectoral muscle removal and mass region extraction is proposed. To find the latent feature, a block of hybrid features is developed and forwarded towards different classifiers. This proposed CAD system produces the recent state-of-the-art (segmentation and classification) results for two benchmark datasets: MIAS and DDSM. It achieved the highest score in term of sensitivity rate 98.40%, specificity rate 97.00% and 97.7% accuracy on MIAS dataset. In fourth methodology, a multi-classification model is presented. Usually, binary classification is performed for classification of breast tumor where it is categorized as an abnormal and normal class or malignant and benign class. There is no mechanism which could further classify or identify each category of abnormal class in order to make diagnostic procedure more reliable for accurate and timely treatment. In this study, the number of descriptors is studied and morphological descriptors are extracted with the help of compound transform on two resolution level of wavelet transform. However, among them, a descriptive set of features is chosen with the help of mRMR which is further used for classification. This method describes absolutely efficient and automated method for multiple kinds of abnormal breast mass classification with an accuracy 94.71%. Fifth technique presented in this thesis is named as a novel segmentation technique based on curve stitching and adaptive hysteresis thresholding. Main idea governing this scheme is to improve threshold-based segmentation algorithm to create an adaptive threshold and apposite templates for preserving salient features about suspicious regions and provide aid in classification. First, a spline-based curve fitting is applied on edges of the breast parenchyma and fill the region with a very low-intensity value. Then it maps on to the original image to isolate and preserve original intensity of breast region free of the pectoral muscle. The proposed method attains the highest sensitivity rate of 96.6% for the MIAS dataset and 96.4% for the DDSM dataset as compared to existing methods. 5. Conclusion and Future Work Chapter 5 In the last methodology, an efficient mammogram classification model is presented for improving the performance of computer-aided diagnosis system. This method assists the radiologists in decision making. In this model a feature matrix is derived from mammogram by using GLCM and HAT transform. Relevant features are selected from the set of descriptors by performing both F-test and T-test. It is analyzed that the feature set selected by performing T-test obtains better results with higher accuracy in BPNN as compared to that of F-test. The simulation was performed by utilizing DDSM and MIAS databases which proved effectiveness of the presented model. It is analyzed in comparison of the listed schemes by simulating them in same platform in respect of AUC and accuracy of ROC. Our analysis reveals that proposed system shows better performance than other systems. In MIAS, an accuracy of 95.0%, for benign-malignant and that of 98.5% for normal-abnormal has been calculated. The similar parameters 98.0%, and 99.0% are obtained in the DDSM database. The investigation conducted in this research work yielded the following important conclusions: ∙ Mammographic images are textured images and better analyzed in the context of multiresolution analysis as it make better distinction between diverse textures present in the mammogram. ∙ The use of a combination of discrete differentiation operator and convex hull algorithm for extraction of any shape and size of breast body can improved, segmentation of accurate breast profile that is unavailable without using this combination. ∙ Mesh of transformation techniques highlights obscured salient descriptors which can improve the performance of the classification model. ∙ Transformation of grayscale image into colored images improve the visual quality of mammograms and achieved high accuracy in splitting regions with different temperature which is further used in screening applications and can reduce false positive rate. ∙ Hybrid features or different kinds of features can better discriminate the breast tissue which improves classification of the breast mass. ∙ Mesh of features extraction at the different level of transformation can also enhance the visual and textural properties of breast mammograms which further yields to reduce false decisions. ∙ Calculation of GLCM at different angels from the transformed image can also improve the detection rate. ∙ Analysis of mammogram texture using different view angle on both TOP-Hat and Bot Hat transformations can capture deep significant details which may cause improve classification accuracy at initial stage of tumor. Timely diagnosis of breast cancer helps to reduce the mortality rate among women.