Improving Deep Learning Classification of Imbalanced Medical Dataset using GANs

Abstract

Computer Vision is one of the latest topics for research. Semantic segmenta tion, classification and localization, object detection, and instance segmen tation are major tasks of computer vision. These tasks are performed by well-designed deep neural networks. Convolutional Neural Networks are de signed for image classification tasks. Classification will not give the desired accuracy if the dataset used is imbalanced. Imbalanced datasets result in biased classification and overfitting, making CNN architectures vulnerable to misclassification. Medical datasets are not easy to collect and preprocess. It has created a problem to automate the health system. In recent years, Generative Adversarial Networks (GANs) have gained remarkable attention from the research community due to their ability to learn the underlying distribution of complex real-world image data. Till now a variety of GANs are introduced. In this study, a novel GAN-based architecture is proposed to improve the classification accuracy of deep learning models using COVID 19, Normal, and Pneumonia images as datasets. Dense-MFC-GAN is used to generate minority class samples. These generated samples are then added to the base dataset for evaluation purposes. This augmented dataset outper forms the performance metrics used to