Application of Adaptive Resonance Theory Neural Network for MR Brain Tumor Image Classification

Introduction

Computer-assisted surgical planning and advanced image-guided technology have become increasingly used in neurosurgery (Webster, 1998). Modern medical imaging technology such as Magnetic Resonance Imaging (MRI) has given physicians a noninvasive means to visualize internal anatomical structures and diagnose a variety of diseases (Haack et al., 1999). MR images are typically interpreted visually and qualitatively by radiologists. The need for quantitative information is becoming increasingly important in clinical and surgical environment. Classification (Norbert et al., 2008) of brain tumor from magnetic resonance (MR) images is a vital process for quantitative follow-up assessment and treatment planning. Brain tumor classification is also essential for understanding the differences of healthy subjects and subjects with tumor. A detailed analysis on brain tumors and edema is reported in many medical journals. A recent work clearly illustrates the necessity for tumor and edema segmentation (Xie et al., 2009).

Automating the classification process is a challenging task. Besides being automated, the technique should be accurate and robust. Several computer assisted methods have been proposed for the classification and quantification of brain tumors. Linear Discriminant Analysis (LDA) classifiers (Theresa et al., 2001), Principal Component Analysis (PCA) (Suma & Murali, 2007) are some of the techniques used for brain tumor classification problem. But, due to their linear boundary, overlapping classes are very difficult to handle. Brain tumor classification has been performed using long echo proton MRS signals (Huffel, 2004). The major limitation is the limited number of available spectra for the tumor types which results in inferior classification accuracy. Brain tumor classification has also been implemented using wavelets (Lee & Hsiao, 2000). But, it suffers from low convergence rate. Expectation-maximization techniques are also used for brain tumor classification (Pohl et al., 2004). But the major drawback is the requirement of a spatial probabilistic atlas that contains expert prior knowledge about the brain structures.

Statistical classifiers, Probabilistic classifiers, Artificial Neural Networks (ANN) are some of the widely used image classifiers (Rafael & Woods, 2002). The major drawback of the statistical classifiers is its inability to classify accurately. On the other hand, probabilistic classifiers suffer from the setback of difficulty in estimating the conditional probabilities. But ANNs outperform the other classifiers because of its flexibility, scalability, tolerance to faults, accuracy, learning, etc (Freeman & Skapura, 2002). Thus, artificial neural networks are highly preferred over the other classification techniques.

The application of ANN for brain tumor classification has been widely analyzed by the researchers. Zumray et al. (1996) elaborates the inferior results of multilayer perceptron for the biomedical image classification problem. The Self Organizing Feature Map (SOFM) ANN based algorithms (Yan & Zheru, 2005) shows excellent results in the classification of brain tumor images. Other studies (Anagnostopoulos et al., 2001) based on learning vector quantization (LVQ) ANN show the potential of these architectures in the case of supervised classification and proved more convenient than traditional ANN approaches such as back-propagation ANN. Hopfield neural networks (HNN) (Sammouda et al., 1994) also prove to be efficient for unsupervised pattern classification of medical images, particularly in the detection of abnormal tissues. The use of ART2 network for pattern recognition has been studied by Solis et al. (2001). In any case, it is of primary importance to establish methods to select the features used as input for these networks.