In-depth analysis of deep learning architectures for brain tumor classification in MRI scans

Abstract

One of the deadliest and most difficult tumors to cure is a brain tumor. Patients diagnosed with brain tumors tend to have a comparatively shorter lifespan. This tumor can affect any individual of any age. To mitigate the damages of brain tumors, early prognosis, and diagnosis are mandatory for a comparatively longer lifespan. Our primary goal is to develop a functional convolutional neural network (CNN) model that can reliably identify brain tumor cells in a patient’s magnetic resonance imaging (MRI). Unfortunately, this is a hard task as there are not many resources available as around 2 to 3 cases occur each year in 100,000 individuals in Bangladesh. For this purpose, a dataset was collected and augmented into a larger dataset by splitting, rotating, changing orientation, etc. Three categories were added to the dataset: training, validation, and testing where 70% of the data was for training, 15% for validation, and 15% for testing. Finally, we trained our dataset for 50 epochs to get the accuracy rate and then tested the same data sets with other pre-trained models like MobileNetV2, DenseNet121, and ResNet50. In this course of training our custom CNN model, we gained the highest accuracy rate, which is 97.07% in training, 95.99% in validation, and 96.51% for testing.