Brain Tumor Classification on Magnetic Resonance Imaging Images using Convolutional Neural Network with Cycle Generative Adversarial Network and Extreme Gradient Boosting

Abstract

Abstract. With the current advancement in technology, image classification process can be carried out through computer processing. This can also be applied to various fields, one of which is the health sector. The health sector is known for its high complexity in pattern recognizing of diseases. One of the diseases that is difficult to classify is brain tumors.            

Purpose: This study aims to improve the accuracy of classification in brain MRI images, which are known to have a small and unbalanced sample. This limitation poses challenges in developing an effective classification model. The classification model is highly dependent on the quantity of data used for training. Therefore, data augmentation techniques play a crucial role in influencing the model's performance.

Methods/Study design/approach: In this study, CNN model using VGG-19 architecture was used to learn feature of brain tumor in brain MRI images. Additionally, CycleGAN is used to augment and balance the data, addressing issues related to data scarcity and imbalance, thus improving diversity of the dataset. And then, XGBoost is applied to classify the feature learned by the CNN model.

Result/Findings: CycleGAN has the ability to generate new image by transferring characteristics between images with different classes, making it a suitable to replace traditional data augmentation techniques in CNN. Additionally, XGBoost can be used to improve the classification results by classifying the features learned by CNN model during the training process. The proposed combination method achieves a highest accuracy of 97.37%.

Novelty/Originality/Value: CNN combined with CycleGAN and XGBoost successfully improved the accuracy of the model and addresed data scarcity and imbalances in the dataset used. This combined method can improve the accuracy of the classification models. This is proven by an accuracy increase of 0.36% when compared to previous research.