In this paper, we propose an efficient fusion framework for brain magnetic resonance (MR) image classification using deep learning and handcrafted feature extraction methods, namely histogram of oriented gradients (HOG) and local binary patterns (LBP). The proposed framework aims to: (1) determine the optimal handcrafted features by Genetic Algorithm (GA) (2) discover the fully connected (FC) layers features using fine-tuned convolutional neural network (CNN) (3) employs the canonical correlation analysis (CCA) and the discriminant correlation analysis (DCA) methods in feature level fusion. Extensive experiments were conducted and demonstrated the classification performance on three benchmark datasets, viz., RD-DB1, TCIA-IXI-DB2 and TWB-HM-DB3. The mean accuracy of 68.69%, 90.35%, and 93.15% from CCA and 77.22%, 100.00%, and 99.40% from DCA was achieved by the Support Vector Machines (SVM) sigmoid kernel classifier on RD-DB1, TCIA-IXI-DB2, and TWB-HM-DB3 respectively. The obtained results of the proposed framework outperform when compared with other state-of-art works.
Key words: Deep learning features, Handcrafted features, Canonical correlation analysis, Discriminant correlation analysis, Support vector machines.
|
