Comparison of classifiers for brain tumor segmentation

Abstract

Nowadays numerous efforts and promising results are obtained in medical imaging processing, although reproducible segmentation and classification of tumors is still a challenging task. The difficulty consists of the different shapes, locations and image intensities of these tissues. In this article we present our discriminative segmentation system for brain tumor delimitation from multimodal MR images. The detection of tumor requires a well-defined process-sequence on every analyzed MRI including preprocessing, feature extraction, classification and post-processing. The discriminative models are trained from annotated image databases and build their decision function around a classifier algorithm. In machine learning there are a lot of advanced classifiers that can be used for segmentation task. The choice of the most adequate classifiers is not straightforward. The SVM, the AdaBoost and the Random Forest (RF) are among the most 10 best classifiers and are often used in image segmentation. The goal of this paper is to analyze and compare these three classification techniques and their obtainable performances or the same segmentation framework. First, we present our framework for brain tumor segmentation. In the theoretical part we briefly present each classifier, emphasizing the advantages and disadvantages of them. In practice, we trained all these three classifiers on the same data set and tested them on 10 image sets, which were not used during training phase. The segmentation performance was evaluated with the Dice coefficients, computing them in each test case separately. Finally, we statistically compared the provided results. At the end, we evaluated the dependence of segmentation accuracy on training set size, providing practical information for possible users of the classifiers created.