A Fine Tune VGG16 Model Based on Ablation Study for Diagnosing Brain Tumor from MRI Images

Abstract

Brain tumor recognition by magnetic resonance imaging (MRI) is crucial because it improves survival rates and allows them to plan treatments accordingly. The patient is at risk if a tumour in the brain, which is made up of a cluster of abnormal cells, spreads to nearby tissues. MRI is the primary technique of imaging which is used for determining the extent of brain tumours. Deep Learning techniques have rapidly expanded in popularity in computer vision applications due to the abundance of data available for training models and advancements in designing models that provide more accurate estimations. When using deep learning techniques to recognize and categorize brain tumors, magnetic resonance imaging (MRI) has produced satisfactory performance. In this paper, we develop a strong deep-learning model which classifies brain tumors into four groups depending on MRI scans using a CNN. Unsolicited areas of brain tumours are deleted with the help of artefact removal, lowering noise, and quality-enhanced images. With improved image quality the cancer is tinted. The number of MRI images has increased using two augmentation techniques. The augmented dataset was analyzed by a number of CNN architectures, including VGG19, MobileNetV2, InceptionV3, VGG16, and Mobile Net. In this situation, VGG-16 offers the highest level of accuracy. The best model was then chosen, and a ablation study was performed on it based on the hyperparameters. The best outcomes were achieved by the hyper-tuned VGG16, which had test accuracy of 98.56% and validation and test accuracy of 99.23%.