Classification of Brain Abnormalities Expressed in MRI Images Using HOG Feature Descriptors and Neural Networks

An approach aimed at detecting types of brain abnormalities in MRI images is presented, which provides high performance and simplicity of design. The developed model proposes to extract handcrafted numerical features from MRI images using a fine-tuned histogram of oriented gradients algorithm. The obtained features undergo dimensionality optimization by applying principal component analysis before being passed to a 6-layer custom neural network that is trained to perform the tumor identification task. In addition, our work investigates the feasibility of performing augmentation in the feature space using the synthetic minority oversampling technique. The proposed approach achieved a high accuracy of 99.65 % and an F1-score of 99.64 %, superior to many works in recent literature.