DCGAN-based data augmentation for improving CNN performance in meningioma classification

Abstract

Using Magnetic Resonance Imaging (MRI) to scan the brain and using the image to identify whether the patient has a brain tumor or not is the common way doctors use it today. However, as this may potentially add to the workload of healthcare professionals, it becomes crucial to explore methods for automating image identification. One effective algorithm for this purpose is the utilization of a Convolutional Neural Network (CNN) network. However, when applying a CNN network to discern whether an individual has meningioma or not, it becomes evident that the available data may be limited. Meningioma is relatively uncommon, and not all associated images have been made accessible for analysis. The shortage of original samples makes it hard to train the CNN network and has relatively low accuracy. In this case, this study tries to use DCGAN to generate more images based on the original sample. By comparing the accuracy and f1 score of the CNN network, this study finds that implementing images has improved the performance of the CNN network. By implementing the images, the DCGAN generates, the accuracy for the same CNN network to identify whether the images have meningioma or not is increased from 93.53 percent to 97.75 percent. The f1 score also increased from about 0.9187 to about 0.9738.