Generative Adversarial-based Framework for Classification using Imbalance Data: Application to Pneumonia Detection in Chest Radiographs

Pneumonia is an infectious lung illness in humans that affects one or both lungs which is caused by the fungus, bacteria, viruses, and other microorganisms. Due to the difficulty of identifying this lung illness at early stages, children below the age of five years and aged people are easily affected by this illness. In this light, early diagnosis and treatment of these diseases are critical. Pneumonia detection requires experienced specialists to identify the pneumonia symptoms from chest X-ray radiography, which is time-consuming and costly in clinical practice. Thus, an automated pneumonia classification from X-ray chest data is highly valuable. Despite the extraordinary performance of the deep learning approaches, some of the proposed methods suffer from issue of imbalance data, insufficient number of labeled data for a specific disease type, as well as errors in the labelled data. In this light, we set out to propose a method with the least possible supervising for training with the minimum amount of data¸ to provide a robust solution for the issue of training with imbalance dataset. The proposed framework tends to regenerate/estimate new features (data-augmentation in the feature map) from X-ray images with GAN. Features are produced by a competition between the Generator, Feature Extractor network, and the Discriminator network. The novel Semi-Supervised GAN-Based method to overcome the issue imbalance data was evaluated using an imbalance dataset for pneumonia detection. The results showed that the proposed framework outperforms the baseline Network, that used an end-to-end supervised learning approach, with a 5.3% improvement in classification accuracy.