Classification of Brain Malignant Tumors Using MRI Scans and CNN Architectures With Optimized Hyperparameters

Abstract

Abstract Based on Magnetic Resonance Imaging (MRI) scans and Convolutional Neural Network (CNN) architectures with optimized hyperparameters, this study proposes a machine learning system for classification of brain malignant tumors. Using a two-level approach, the first level identifies the most suitable MRI modalities for classification by evaluating the performance of the modalities under different hyperparameters. The second level consists of the development and comparison of two CNN architectures, Inceptionv3 and DenseNet121, based on their respective training methods. As part of the optimization process, various hyperparameters are tested, such as epochs, batch sizes, learning rates, and dropout rates. Additionally, it was found that learning rate has an important impact on the training phase with a lower learning rate resulting in a smoother, more consistent process. As a result of optimizing the system, it has been demonstrated that it can accurately classify brain malignant tumors, particularly Glioblastomas.