Comparative Study of Optimization Algorithms in CNNs for Brain MRI Image Classification

Abstract

Brain MRI often reveals long-standing diseases of the nervous system, such as multiple sclerosis, dementia, a stroke, and brain malignancies. In addition to that, the most accurate method of brain MRI, besides the diagnosis of pituitary gland diseases, is the method diagnosing the vessels of the brain and eyes and the organs of the inner ear. On the other hand, many methods of loading medical pictures have been developed with brain MRI data, often to diagnose diseases and monitor health via it. Convolutional neural networks belong to deep learning and are widely used for input from the visual domain. The most common use of CNN is in natural language processing and recommendation systems, image classification, medical imaging, and image and video recognition. This work is divided into several parts. The Msoud dataset, used in this study, consists of 7023 MRI images, which were made by the Fighshare, SARTAJ, and Br35H datasets. The MRI images are of four classes, that is, healthy brains, brains with glioma, brains with meningioma, and pituitary. In this research work, the doing of different pre-processing of the MRI input to make the images ready for the model to be trained is done. The architecture is made up of dense layers such that after each set of convolutional layers, there is a max-pooling. Eventually, batch normalization and dropouts in the training are stabilized to reduce overfitting. The proposed CNN compared with other studies and many transfer learning models found the proposed model to achieve significant accuracy of 99.00%, 98% and 97% for using Adamax, Adam and RMSprop optimizers respectively.