Classifier ensemble generation and selection with multiple feature representations for classification applications in computer-aided detection and diagnosis on mammography

Highlights

• Novel ensemble classifier framework for improved classification of breast lesions.

• Ensemble generation algorithm using different types of breast lesion features.

• Ensemble selection mechanism to find an optimal subset of component classifiers.

• Impressive classification performance by comparing single classifier based methods.

Abstract

This paper presents a novel ensemble classifier framework for improved classification of mammographic lesions in Computer-aided Detection (CADe) and Diagnosis (CADx) systems. Compared to previously developed classification techniques in mammography, the main novelty of proposed method is twofold: (1) the “combined use” of different feature representations (of the same instance) and data resampling to generate more diverse and accurate base classifiers as ensemble members and (2) the incorporation of a novel “ensemble selection” mechanism to further maximize the overall classification performance. In addition, as opposed to conventional ensemble learning, our proposed ensemble framework has the advantage of working well with both weak and strong classifiers, extensively used in mammography CADe and/or CADx systems. Extensive experiments have been performed using benchmark mammogram dataset to test the proposed method on two classification applications: (1) false-positive (FP) reduction using classification between masses and normal tissues, and (2) diagnosis using classification between malignant and benign masses. Results showed promising results that the proposed method (area under the ROC curve (AUC) of 0.932 and 0.878, each obtained for the aforementioned two classification applications, respectively) impressively outperforms (by an order of magnitude) the most commonly used single neural network (AUC = 0.819 and AUC =0.754) and support vector machine (AUC = 0.849 and AUC = 0.773) based classification approaches. In addition, the feasibility of our method has been successfully demonstrated by comparing other state-of-the-art ensemble classification techniques such as Gentle AdaBoost and Random Forest learning algorithms.

Introduction

Breast cancer is the most common form of cancer among women and is the second leading cause of death (Kopans, 2007). To reduce the workload of radiologists and to improve the specificity and sensitivity in detection of breast cancer, two different types of automated screening systems are being developed (Suri & Rangayyan, 2006): (1) Computer-aided Detection (CADe) and (2) Computer-aided Diagnosis (CADx). Table 1 provides a brief review of CADe and CADx systems. Current CADe and/or CADx systems have been clearly shown to be quite sensitive in its ability to detect cancer, but one of their main drawbacks is the high number of FPs (defined in Table 1) (Suri and Rangayyan, 2006, Sampat, 2005). Hence, high FP rate for mass detection and diagnosis remains to be one of the major problems to be resolved in CADe/CADx study (Suri and Rangayyan, 2006, Sampat, 2005, Tang et al., 2009).

In typical CADe (or CADx) systems, classifier design is one of the key steps for determining FP rates (Suri and Rangayyan, 2006, Sampat, 2005). Thus far, research efforts have mostly been focused on the design of the single classifier in both CADe and CADx systems (Suri and Rangayyan, 2006, Sampat, 2005, Tang et al., 2009, Chan et al., 1999). It should be noted that there are two critical limitations within the classifier design process in mammogram images. First, the large variability in the appearance of mass patterns (Cheng et al., 2005, Velikova et al., 2013) – due to its irregular size, obscured borders, and complex mixtures of margin types – makes classification task quite difficult. Second, research in mammography is characterized by a restricted training data, due to cost, time, and availability to patient medical information and mammography images (Suri and Rangayyan, 2006, Bilska-Wolak and Floyd, 2004). On the other hand, the number of available features (arising from the integration of multiple heterogeneous feature types) is large (Cheng et al., 2005, Jesneck et al., 2006, Wei et al., 1997) (typically, in the thousands) relative to the number of training samples, so-called curse of dimensionality (Kuncheva, 2004). For these reasons, a single classifier design may face a great challenge in achieving a level of FP reduction that meets the requirement of clinical applications.

In this paper, to overcome the aforementioned limitations, we propose a new and novel ensemble classifier framework for classification applications (explained in Table 1) in mammographic CADe and CADx. This paper improves and extends preliminary work presented in Choi, Kim, Plataniotis, and Ro, (2012). In particular, this paper presents a new ensemble selection approach for selecting an optimal subset of base classifiers, aiming to further improve generalized (testing) classification performances. An improved ensemble generation technique is also outlined in the paper by introducing an advanced mechanism that allows the use of strong classifiers extensively used in mammography computer-aided detection and diagnosis systems. In addition, more insightful discussion of our ensemble generation on the local learner hypothesis viewpoint is provided. Moreover, we report integrated experimental results that are more extensive and rigorous in the following aspects: (1) additional assessment of our proposed ensemble classification on computer-aided diagnosis application; (2) the comparison of other state-of-the-art ensemble classification techniques; (3) comprehensive analysis using more classifier models.

The contents of the paper are organized as follows: Section 2 reviews previous work on classification of breast masses on mammograms in CADe and CADx systems. In Section 3, the region-of-interests (ROIs) segmentation and feature extraction methods used in our study are briefly described. Section 4 explains in detail the proposed ensemble classification framework. Section 5 contains the details of the image databases, and experimental setup and condition. In Section 6, we present a series of experimental results to demonstrate the effectiveness of the proposed method. Finally, concluding remarks are provided in Section 7.