Effects of Filter on the Classification of Brain MRI Image using Convolutional Neural Network

Abstract

Brain MRI image produces a clear anatomical view of brain and any small abnormality of brain is perceptible by MRI image. Brain tumor classification, detection, and segmentation are huge concern of researchers. Among various machine-learning algorithms such as -K-Nearest Neighbour, Support Vector Machine, Artificial Neural Network, and Convolutional Neural Network (CNN); CNN acquired better position in image classification. CNN classification accuracy depends on some network parameter as convolutional filters, rectification functions, polling functions, and iteration numbers etc. It is a problem to determine effective values of convolution filter size, number and convolution stride for better accuracy in classification. CNN has capabilities to produce and learn effective features on large datasets. We have provided some directions to determine effective values for CNN filter size and number in Brain MRI image classification. Our research work’s findings are - square filters are better than rectangular filter; accuracy of filter size 3*3 to 10*10 higher than larger filter size; larger number of filters produce more complexities with higher classification accuracy; increasing of convolutional stride results decrease in network accuracy. This work will be helpful and may be a direction for researchers who are interested to work with Convolutional Neural Network in image classification fields.