Differential diagnosis of CT focal liver lesions using texture features, feature selection and ensemble driven classifiers

Summary

Objectives

The aim of the present study is to define an optimally performing computer-aided diagnosis (CAD) architecture for the classification of liver tissue from non-enhanced computed tomography (CT) images into normal liver (C1), hepatic cyst (C2), hemangioma (C3), and hepatocellular carcinoma (C4). To this end, various CAD architectures, based on texture features and ensembles of classifiers (ECs), are comparatively assessed.

Materials and methods

Number of regions of interests (ROIs) corresponding to C1–C4 have been defined by experienced radiologists in non-enhanced liver CT images. For each ROI, five distinct sets of texture features were extracted using first order statistics, spatial gray level dependence matrix, gray level difference method, Laws’ texture energy measures, and fractal dimension measurements. Two different ECs were constructed and compared. The first one consists of five multilayer perceptron neural networks (NNs), each using as input one of the computed texture feature sets or its reduced version after genetic algorithm-based feature selection. The second EC comprised five different primary classifiers, namely one multilayer perceptron NN, one probabilistic NN, and three k-nearest neighbor classifiers, each fed with the combination of the five texture feature sets or their reduced versions. The final decision of each EC was extracted by using appropriate voting schemes, while bootstrap re-sampling was utilized in order to estimate the generalization ability of the CAD architectures based on the available relatively small-sized data set.

Results

The best mean classification accuracy (84.96%) is achieved by the second EC using a fused feature set, and the weighted voting scheme. The fused feature set was obtained after appropriate feature selection applied to specific subsets of the original feature set.

Conclusions

The comparative assessment of the various CAD architectures shows that combining three types of classifiers with a voting scheme, fed with identical feature sets obtained after appropriate feature selection and fusion, may result in an accurate system able to assist differential diagnosis of focal liver lesions from non-enhanced CT images.

Introduction

One of the most common and robust imaging techniques for the detection of hepatic lesions is computed tomography (CT) [1]. Although the quality of CT images has been significantly improved during the last years, it is difficult in some cases, even for experienced doctors, to make a 100% accurate diagnosis. In these cases, the diagnosis has to be confirmed by administration of contrast agents, which is related with renal toxicity and allergic reactions, or invasive procedures (biopsies). During the last years, along with the developments in image processing and artificial intelligence, computer-aided diagnosis (CAD) systems, aiming at the characterization of liver tissue, attract much attention, since they can provide diagnostic assistance to clinicians, and contribute to reduction of the number of required biopsies.

Various approaches, most of them using ultrasound B-scan and CT images, have been proposed based on different image characteristics, such as texture features, estimated from first- and second-order gray level statistics, and fractal dimension estimators combined with various classifiers [2], [3], [4], [5]. Texture analysis of liver CT images based on spatial gray level dependence matrix (SGLDM), gray level run length method (GLRLM), and gray level difference method (GLDM) has been proposed by Mir et al. [6], in order to discriminate normal from malignant hepatic tissue. Chen et al. [7] have applied SGLDM texture features to a probabilistic neural network (P-NN) for the characterization of hepatic tissue (hepatoma and hemangioma) from CT images. Additionally, SGLDM-based texture features fed to a system of three sequentially placed neural networks (NNs) have been used by Gletsos et al. [8] for the classification of hepatic tissue into four categories. Although a lot of effort has been devoted to liver tissue characterization, the developed systems are most of the times limited to two or three classes of liver tissue and/or do not gain from the interaction of different texture characterization methods, or the combination of different classifiers.

In order to select the most robust characteristics from an initial high-dimensional feature set, that might be derived from different feature extraction techniques, feature selection methods can be applied. Deterministic, or stochastic feature selection methods decrease the feature extraction costs of the classification system, and may also enhance its performance [9]. During the last years, an increasing number of researchers are using genetic algorithms (GAs) for dimensionality reduction. The use of GAs for feature selection was first introduced in 1989 [10], and since then, GAs have been successfully applied to a broad spectrum of dimensionality reduction studies [11], [12], [13]. According to Ref. [8], the use of GAs results in more robust feature vectors as compared to other deterministic feature selection techniques, in problems related to liver tissue classification from CT images.

In the last decade, the use of multiple classifier systems has been proposed in order to optimize the performance of CAD systems. A set of classifiers whose individual predictions are fused through a combining strategy, usually a voting scheme, to classify new examples constitutes an ensemble of classifiers (EC) [14]. The attraction that this topic exerts on machine learning and diagnostic decision support research is based on the premise that ECs are often much more accurate than any individual classifier of the set [15], [16]. Early diagnosis of melanoma has been facilitated by combining three types of classifiers, namely linear discriminant analysis (LDA), k-nearest neighbor (k-NN), and a decision tree, with a voting scheme [17]. A multiple classification system based on a committee of NNs, trained by the Levenberg–Marquardt algorithm, along with a voting scheme across the NN outputs has been used by Jerebko et al. [18] for the detection of colonic polyps in CT colonography data. Furthermore, a novel system for diabetes diagnosis has been proposed [19], which is based on retinal images fractal characteristics and applies a voting scheme across the outputs of an EC consisting of a back-propagation trained NN, a radial basis function NN, and a GA-based classifier. The use of texture features and shape parameters along with a multi-classifier modular architecture composed from a self-organizing map (SOM) and/or k-NN classifiers has been recently proposed by Christodoulou et al. [15], aiming at the characterization of carotid plaques for the identification of individuals with asymptomatic carotid stenosis at risk of stroke.

Although previous studies [20], [21] have shown that CAD systems based on various texture features and ECs can enhance the diagnosis efficiency of CT focal liver lesions, the evaluation of the proposed methods on small-sized samples constitutes a significant drawback for these studies. To address this drawback, re-sampling methods like cross-validation, jack-knife, and bootstrap can be applied [22], [23], [24]. In Ref. [18], cross-validation has been applied for sensitivity estimation of a colonic polyp detection system, while in Ref. [25] the bootstrap method has been used for the development of a diagnosis system able to differentiate benign and malignant tumors from breast ultrasound images. The bootstrap method, which was introduced by Efron [22] as an approach to calculate confidence intervals for parameters where standard methods cannot be applied, is based on re-sampling with replacement. The comparative assessment of various re-sampling methods has shown that the bootstrap method provides less biased and more consistent results than the jack-knife method [26].

The principal aim of the present paper is to assess the potential of ECs in the development of a CAD system able to discriminate four hepatic tissue types (normal liver, hepatic cyst, hemangioma, and hepatocellular carcinoma) from non-enhanced CT images. Furthermore, the use of a variety of texture features as input to the CAD system is examined while the application of feature selection based on a GA is investigated aiming at improving the resulting classification performance. In order to overcome problems with small data sets and biased classification performances, the bootstrap method is applied. In this framework, five different CAD architectures based on the above design concepts are comparatively assessed.

The rest of the paper is organized as follows: in Section 2, the generic system design concepts are presented, including description of the data used, the methodology of feature extraction and selection, the ECs and the applied voting schemes, as well as the five alternative architectures of the CAD system. In Section 3, the experimental results of the five CAD system architectures are presented and compared, followed by conclusions presented in Section 4.