Stacked Sparse Autoencoder and Softmax Classifier Framework to Classify MRI of Brain Tumor Images
Authors: Ahmed Aldhahab, Sarmad Ibrahim, Wasfy B. Mikhael
Number of views: 211
Classification of a brain tumor is a critical step in the design of computer-aided diagnosis systems for Magnetic Resonance Image (MRI) analysis. This work presents an efficient algorithm to classify a tumor in brain MRI images using statistical-based features and deep neural network. Data, within the region of interest, is transformed into two-dimensional discrete Gabor filter and wavelet transform. These filters are combined in this algorithm as directional transformation methods for utilizing all information in all orientations of the MRI input image. MRI Features are extracted based on the first and second order statistics from both domains. Two types of neural network classifiers are employed: Stacked Sparse Autoencoder (SSA) and Softmax Classifier (SMC). Two regularization functions are used in the training of the SA, sparsity regularization and L2-weight regularization. Sparsity regularization controls the firing of the neurons in the hidden layer, whereas L2-weight regularization reduces the effect of the overfitting and improves the performance of the SA. Two datasets are used to evaluate the proposed algorithm. The first dataset consists of 3,064 of T1-weighted MRI slices with three kinds of tumors: Pituitary, Glioma, and Meningioma. The second dataset consists of 200 MRI slices with low-grade and high-grade Glioma tumor collected from the BRATS dataset. The performance of the proposed algorithm is validated using the experimental results in terms of accuracy, specificity, and sensitivity compared to the existing algorithms. For the first dataset, the accuracy obtained is 94.0%, the sensitivity of Meningioma, Glioma, and Pituitary is 87.44%, 97.29%, and 94.27%, respectively, and the specificity of Meningioma, Glioma, and Pituitary is 98%, 96.89%, and 96.78%, respectively. For the BRATS dataset, the accuracy, the specificity, and the sensitivity achieved are 98.8%, 100%, and 100%, respectively.