Mammogram Analysis: Tumor Classification Literature Survey Report Geethapriya Raghavan [email protected] EE 381K - Multidimensional Digital Signal Processing Spring 2005 Abstract Breast cancer is the main causes of cancer related mortality among American women. The use of screening mammography as the most reliable method for early diagnosis of breast cancer is widely recommended with the introduction of several Computer Aided diagnosis (CAD) techniques. I am analyzing some of the pattern recognition techniques that have been most effective in classifying tumor as benign and malignant – support vector machine (SVM), kernel Fisher discriminant (KFD), relevance vector machine (RVM) and a multiresolution pattern recognition method using wavelet transform. These methods have been developed and implemented in statistical learning theory over the past decade and are expected to give promising classification results for efficient tumor diagnosis.1 Introduction Breast cancer is among the most common and deadly of all cancers, occurring in nearly one in ten women. Mammography is a uniquely important type of medical imaging used to screen for breast cancer. All women at risk go through mammography screening procedures for early detection and diagnosis of tumor. Currently there are no methods for breast cancer, which is why early detection becomes important to achieve high survival rates. A typical mammogram is an intensity x-ray image with gray levels showing levels of contrast inside the breast which characterize normal tissue and different calcification and masses. The contrast level of a typical mammogram image is proportional to the difference in x-ray attenuation between different tissues. In general, a clear separation between normal functioning tissue and abnormal cancerous tissues is difficult to identify since their attenuation is very similar. Important visual clues of breast cancer include preliminary signs of masses and calcification clusters. A mass is a localized collection of tissue seen in two different projections and calcifications are small calcium deposits. Masses and calcium deposits are easy to see by x-ray because they are much denser (highly attenuate x-ray) than all other types of soft tissues around. Unusually smaller and clustered calcifications are associated with malignancy while there are other calcifications (diffuse, regional, segmental and linear) that are typically benign. Such calcifications are termed as microcalcifications. In mammogram analysis, the focus is mainly on two types of lesions – masses and microcalcifications (Fig.1a and 1b). In theearly stages of breast cancer, these signs are subtle making diagnosis by visual inspection difficult. With millions undergoing mammography procedures, the need for quick and reliable computer based tools is strongly felt. Fig 1(a) Fig 1(b) Fig1. (a) Mammogram showing a big mass and (b) a clustered microcalcification 2 Background Intensive research work has been undertaken in the development of automated image analysis methods to assist radiologists in the identification of abnormalities. The role computers play in mammogram analysis is threefold: detection, diagnosis and noise cancellation. Detection involves identifying cancerous tissues in a mammogram. Early detection of breast cancer by mammography depends on the production of excellent images and competent interpretation. Mammography alone cannot prove that a suspiciousarea is malignant or benign. To decide that, the tissue has to be removed for examination using breast biopsy techniques. A false positive detection may cause an unnecessary biopsy. Some of the more important pitfalls encountered two decades ago with low contrast and poor image quality in mammography are presented in [1]. Diagnosis using mammograms is aimed at classifying the detected cancerous regions as benign or malignant. A review of several studies demonstrating how CAD tools help in tumor diagnosis is presented in [2]. 3 Diagnosis Tools The diagnosis task is modeled as a two-class classification task. Features are extracted from Regions of Interest (ROIs - the region containing the masses or the microcalcification) containing the abnormality (the training phase) and each ROI is classified using a classification algorithm (the testing phase). In most cases, the classification algorithm used is a supervised method that is first trained on a set of sample images called the training set. The performance of the algorithm is then tested on a separate testing set. The metrics used to report the accuracy of these algorithms are sensitivity and specificity. Sensitivity is defined as a lesion for which the CAD predicts that it is cancerous and it is actually found to be malignant. Specificity is the fraction of benign lesions that are correctly identified by the CAD as being benign. A plot of sensitivity versus specificity is called a Receiver Operating Characteristic (ROC) curve and this is used to report the performance of the CAD technique used [3]. The main parameter studied is the area under the ROC curve, Az. Higher the value of Az, better isthe performance of the Cad technique used. Range values Az of any technique can take is from 0 to 1. Hence, a good CAD tool has values closer to one. Classifying a mammogram with a cluster of microcalcifications is more challenging than doing the same with masses because of their erratic shapes, size, density and texture. Due to their high success rates [4], I investigate the following contemporary methods. a. Support Vector Machine (SVM) In recent years, SVM learning has found a wide range of real-world applications, including object recognition [5], speaker identification [6] and face detection in images [7]. The formulation of SVM learning is based on the principle of structural risk minimization. Instead of minimizing an objective function based on the training samples [such as mean square error], the SVM attempts to minimize
View Full Document
Unlocking...