Brain tumor detection with convolutional neural network

Abstract

The brain is the command center of our nervous system, which enables thoughts, memories, movements, and emotions. In other words, it is the most important organ in the human body. The human brain is very vulnerable to tumors, as merely growing old can be the cause of a tumor. Furthermore, the effects of a tumor can be fatal to a person because, as the tumor grows inside the brain, it can deform the structure of the brain and cause several diseases, the most fatal being cancer in the brain. Hence, to prevent such severe diseases, early detection of tumors is critical for a patient’s treatment. Moreover, modern technology has emerged to excellent heights, as MRI scans and CT scans can detect brain tumor regions. However, to accurately detect where the tumor is situated, a team of doctors is still needed to this day. Therefore, we have planned to use convolutional neural Networks to develop a faster and inexpensive method to detect tumors from MRI images in the early stages. Moreover, we plan to develop a system where our proposed CNN model will be able to detect tumors as well as identify three types of tumors, which are glioma, meningioma and pituitary tumors. Also, if there are no tumors, the system should be able to detect them too. To develop our proposed model, we have used data pre-processing techniques with a combination of gray scaling, One encoding, and CLAHE. Also, we have used a dataset of 6484 MRI images, segmenting them by testing and training. To compare and analyze our proposed model’s performance, we have tested and trained seven pre-trained models with the same dataset. The models are Vgg16, Vgg19, ResNet50, InceptionV3, DenseNet-121, EfficientNetB0, MobileNet and we received the following testing accuracy accordingly: 93.37%, 92.42%, 75.38%, 91.48%, 94.89%, 23.30% and 96.02%. However, the testing accuracy of our proposed model surpassed all the other pre-trained models, as it gained 98.11% accuracy in testing. In conclusion, we have aimed to build a CNN model that exceeds all the other CNN models in terms of overall performance, which is why we have integrated a sufficient amount of parameters to handle any unfavorable situations; however, the parameters are set in such a way that the overall system does not clutter and remains lightweight.