BrainACGAN: Auxiliary Classifier Generative Adversarial Network for Brain Tumor Images

Application of machine learning in the medical field has considerably improved throughout the past decade due to the ability of autonomous process. Algorithms now assist doctors and physicians to detect and diagnose various diseases with minimal errors. However, obtaining sufficient medical data to develop efficient and generalized algorithms is still challenging due to socio-economic, administrative, and demographic issues. Training algorithms with limited data might introduce biases and mistakes, which can lead to misdiagnosis and perilous treatments. Overcoming data limitation is crucial for effective implementation of deep learning models in medical imaging. To solve these limitations, our study used an advanced image generation technique called auxiliary classifier generative adversarial network (ACGAN) to expand MRI brain tumor image dataset. The dataset is enhanced by creating high-quality synthetic images identical to actual images and then real and synthetic images are combined and fed into a customized convolutional neural network to evaluate the effectiveness of image generation. Results exhibit that average classification accuracy is increased significantly from ${8 8 \%}$ to ${9 9 \%}$ after expanding the dataset. Moreover, a notable improvement in precision, recall and F1 score is observed which indicate the promising impact of ACGAN image generation on our research. This study will contribute to overcome the limitation of small dataset and improve the accuracy of classifying and diagnosing brain tumor using deep learning model.