G-ResNet: Improved ResNet for Brain Tumor Classification

Abstract

The brain tumors, are the most common and aggressive disease, leading to a short life expectancy and much pain. Timely and accurate diagnosis is the key factor in improving the survival rate of patients. The main method of identifying brain tumors is to analyze MR image that provides detailed information about brain structure and anomaly detection in brain tissue. With the rapid development of deep learning, especially the improvement of computer vision technology, automatic brain tumor detection is proposed by using Convolutional Neural Networks (CNN) classification. In this paper, we propose a new model called Global Average Pooling Residual Network (G-ResNet) to classify brain tumor images. The model has the following characteristics: (1) Applying the well-established CNN architecture in the field of deep learning named ResNet34 for the classification task. (2) To reduce the number of parameters and avoid overfitting, we use the global average pooling layer instead of the flattened layer for classification. (3) In order to be able to fuse the low-level and high-level features of the network to improve the classification accuracy, we concatenate the feature vectors of different layers. (4) We define a loss function, which is sum of the interval loss and the cross entropy loss. The total loss increases the penalty for misclassification. In summary, our model achieves the classification accuracy of 95.00%, which is significantly better than the previous models.