Early Detection and Classification of Breast Tumor From Mammography

Abstract

Breast cancer is the major cause of increased death rate, especially in women. According to world health organization, about one million new cases of breast cancer are registered worldwide, while the death rate is about 400,000 women annually. Early detection and timely treatment of breast cancer can reduce this death rate. Mammography is traditional and the most recommended method for diagnosis of breast cancer at its initial stage. It is usually laborious for healthcare specialist to identify small abnormal and hidden masses at initial stages of breast cancer due to poor quality of mammographic images and low sensitivity in dense breast tissue. This impediment is further aggravated by obscurity in those images due to presence of pectoral muscle, labels and artifacts. This problem may cause a high rate of false positives which may lead to unnecessary biopsies and high mortality rate. Therefore it is essential that computer-aided diagnosis (CAD) can identify and correctly locate breast tumor at its initial stage to avoid wrong decisions with their dangerous ramifications. CAD system is a recent advancement in current medical imaging for early detection and classification of breast cancer. CAD systems not only improve quality and sensitivity of mammographic images but also improve diagnostic ability of healthcare specialist. The framework of a CAD system for diagnosing breast cancer has several steps which include image acquisition, preprocessing, segmentation, feature extraction and classification. Existing methods of computer-aided breast cancer detection systems bear inadequacies like high false positive rates and low margins of accuracy, sensitivity and specificity. These handicaps directly contribute to defective pathological assessments leading to avoidable fatality cases of women. It is imperative that performance of CAD systems is improved by overcoming their existing limitations regarding automated breast cancer detection. For this goal, six distinct techniques are introduced in this thesis to address above-mentioned limitations and enhance accuracy of automated system for early detection of breast tumor. In this respect, detection of pectoral muscle is a major challenge due to its overlapping area with fibro-glandular tissue especially in the lower region. In first methodology, an improved technique for removal of pectoral muscle of varying sizes and shapes in mediolateral oblique (MLO) view of mammogram is presented. This approach is based on discrete differentiation operator which is an edge detector and computes an approximation of the gradient of image intensity function. For refinement purpose, convex hull technique is also applied. This method is implemented on 322 mammograms provided by mammographic image analysis society (MIAS) and 24 contrast enhanced mammograms. As compared to earlier techniques our approach caters for a diversity of pectoral muscle geometries even in distorted mammograms. The proposed algorithm shows improved results in terms of mean false positive rate 0.99 and the Hausdorff distance 3.52 mm (which should be lower the better). These executions also illustrate that performance of the intended method has immensely improved from 7.08 mm value of Gabour method to 3.52 mm of our proposed method for removal of the pectoral muscle. Second method describes an efficient technique for accurate segmentation of breast lesion which is often achieved a vigorous initiative in the automated diagnosis system using color features and mathematical morphology. Ho wever, existing segmentation methods cannot isolate concerned region of breast mass with high accuracy, particularly in case of mammogram images that contain diverse textures. A breast mass segmentation technique based on a combination of color space and intensity variation analysis is proposed. In this study, the properties of L*a*b* color space are analyzed with focus on the visual perception of the color components. Pixel features are obtained using color-size histogram for textural analysis as dominant property of the mathematical morphology, which has been performed for accurate segmentation of tumor region. This approach is tested collectively on 722 mammograms provided by MIAS and Digital Database for Screening Mammography (DDSM). Proposed algorithm improved performance of mass segmentation by maintaining the good visual integrity and high accuracy rate of 98.00% on MIAS images and 97% on DDSM images. Early screening of skeptical masses in mammograms may reduce the mortality rate among women due to their timely treatment. This rate can be further reduced by development of an improved computer-aided diagnosis system which can over-all yield more accurate detections and less defective assessments. For this purpose, a third methodology named as bi-model processing algorithm is introduced in this study. According to this algorithm region of interest is divided into two parts, first one is called pre-segmented region and the other is post-segmented region. This system follows the scheme of pre-processing technique of contrast enhancement that can be utilized to segregate and extract desired feature of the given mammogram. In next phase, a hybrid feature block is presented to classify the breast tumor accurately. To validate execution of the proposed method, it is experienced to a database provided by the society of mammographic images. Our experimen tal outcomes on this database exhibit the usefulness and robustness of the proposed method. The proposed CAD system achieves the highest sensitivity (98.40 %), specificity (97.00 %) and accuracy rate (97.7 %) on MIAS dataset. In fourth technique, a novel automated system is presented for classifying multiple breast masses based on mammogram images. Usually, breast masses are categorized in two classes called benign and malignant which are further categorized in seven major classes: welldefined masses, calcification, speculated masses, architectural distortion, asymmetry, ill-defend masses and normal region. Each class of abnormality warrants a strong classification model for its analysis. Mostly, classification accuracy of different abnormalities through automated systems is based on extraction of significant morphological characteristics of desired region. A mesh of transformation schemes on two level wavelets transform of mammographic images is implemented to extract an extensive set of morphological features. From there, informative descriptors with the lowest redundancy are chosen and forwarded towards Naive Bayes classifier for efficient classification of abnormalities. The proposed method deals with multiple kinds of abnormal breast mass classification with 94.71% accuracy. In fifth method, a work regarding the tissue identification in breast body is presented for probing abnormalities detection of particular features during the study of the nonfocused region. Moreover, it is significant to reach optimal measurements of breast parenchyma, breast patchy regions of the mammogram, or breast registration for searching precise abnormalities. Based on this study, a novel segmentation technique using curve stitching and adaptive hysteresis thresholding is presented. The proposed method attains the highest sensitivity rate of 96.6% using MIAS database. In sixth technique, a classification model based on combination of top hat transformation and back propagation neural network (BPNN) is presented. This system significantly diagnose the tumor at initial stage and reduces the rate of false assumptions in medical informatics which improves the classification accuracy of computer-aided diagnosis methods. An accuracy of 99.0% is achieved by the proposed method. The primary goal of this research work is to save the lives by overcoming the limitations present in existing early breast tumor detection systems and provide a robust solution to improve the performance of the current early breast cancer detection system and provide a second opinion for the radiologists. It is notice that the all proposed work outcomes are producing an efficient performance during an experimental evaluation on MIAS and DDSM datasets.