Image fusion for following-up brain tumor evolution

Abstract

This paper presents a feature-selection-based data fusion method to follow up the evolution of brain tumors under therapeutic treatments with multi-spectral MRI data sequences. The fusion of MRI data is proposed to use a feature selection method to choose the most important features to classify tumor tissues and non-tumor tissues. Our system consists of three steps for each MRI examination (one examination per four months): feature selection, SVMbased segmentation, and contour refinement. The training of the tumor is carried out only on the first MRI examination (before the treatment). Six feature selection methods are tested in our system. The quantitative comparisons of the methods' and the expert's manual traces demonstrate the effectiveness of the fusion by the feature selection. Results of following-up three patients over one year show that our system can provide a good tool to evaluate the therapeutic treatment.

A fully automatic extraction of magnetic resonance image features in glioblastoma patients

“…A similar method was used by Fletcher-Heath et al 13 for nonenhancing tumor segmentation. Ruan et al 14 used a support vector machine (SVM) classifier and the features selected based on kernel class separability to segment brain tumor regions. In atlas-based segmentation methods, 15,16 the spatial prior knowledge and probabilistic information provided by atlas are used for brain tumor segmentation.…”

A survey of MRI-based brain tumor segmentation methods

“…SVM was treated as a parametrically kernelbased method to deal with supervised classification problems [98] . SVM has been widely used in the field of brain tumor segmentation [99][100][101][102][103] , mainly owing to its great classification ability.…”

“…A very similar approach based on SVM has been proposed [102] , but this method only segmented one tumor region and used a lower number of modalities. Later, this method was improved by the feature selection with kernel class separability and obtained better results [103] . A multi-kernel based SVM integrated with a feature selection and a fusion process was proposed to segment the brain tumor from multi-sequence MRI images [104] .…”

Automated Glioblastoma Segmentation Based on a Multiparametric Structured Unsupervised Classification

“…Most of these techniques fall into the supervised learning approach. In [ 6 , 7 ] Support Vector Machines (SVM) were applied to multiparametric MR datasets to segment health and pathological tissues, and additionally subcompartiments inside these areas. Jensen et al [ 8 ] applied several neural networks to detect brain tumour invasion.…”