Convolutional neural networks combining multiscale convolution and global padding and its application in brain tumor classification

Brain tumors are one of the leading causes of increased mortality in humans. Prior to treating brain tumors, it is essential to examine brain images to determine the presence of a tumor and to differentiate between specific tumor types, which is a basic task in the treatment process. The conventional approach classifies brain tumors manually, but this may be highly subjective due to variability between observers, resulting in vague, inaccurate and time-consuming classification. Therefore, automatic classification is of great significance in helping doctors curing this kind of disease. To handle these issues, a convolutional neural network structure called Multiscale Convolution Model With Global Padding Strategy (MSCNN) is proposed to classify brain tumor MRI images. Multi-scale convolutional structure and global padding feature extraction strategy are utilized in MSCNN, and global average pooling operation is added to replace the flatten layer for network structure optimization, improving the robustness and generalization of the network. The proposed MSCNN was evaluated by the publicly available CE-MRI Figshare dataset, which containes more than three thousand T1-weighted contrast MRI section images of three kinds of brain tumors including meningioma, glioma, and pituitary gland tumor. Additionally, several well-known machine learning image classification algorithms are compared with MSCNN. Experimental studies show that the classification accuracy of MSCNN is $\mathbf{98.91 \%}$, meaning it provides better classification accuracy and performance.