An Evaluation of Different Deep Feature Extraction Approaches for MRI-Based Brain Tumor Detection

Brain tumor is a leading cause of death globally and Magnetic Resonance Imaging (MRI) is a powerful tool for its diagnosis. However, the ability to extract more representative features for efficient characterization of brain tumors is still a highly targeted topic. This paper aims to study the power of different deep learning approaches such as traditional Convolutional Neural Network (CNN), transfer learning via three well-known models (VGG16, InceptionV3, and ResNet50), and convolutional auto-encoder in extracting deep representative features for efficient characterization of brain tumors. The performance evaluation was done using two public datasets of MRI images and various benchmarking evaluation metrics. Both traditional CNN and convolutional auto-encoder achieved 100% accuracy, with the auto-encoder producing the lowest loss and highly productive with balanced datasets. Transfer learning models achieved an overall performance of 99% in different evaluation metrics and were found to be preferred with unbalanced datasets.