Multi-modal feature fusion and its application to brain tumour classification

In this work the objective is to detect whether the given MRI image of a brain tumour is malignant or benign. Recent literature shows that support vector machines (SVM) exhibit high generalisation ability even trained with small set of training data. Initially the images are preprocessed, segmented, different types of features like discrete wavelet transform (DWT) or gist features are extracted from the given MR images, followed by SVM-based classification. Motivated from the fusion-based classification models, in this work we have extracted different representations from the given MR images and fused them to represent the image as a single feature vector. We have applied different fusion techniques to improve the performance. Our experimental studies on bench mark datasets show that fusion techniques can enhance the accuracy of SVM classification for brain tumour classification. Along with fusion we analyse the efficiency of various kernels on the classifier's performance.