Generative Adversarial Network based Chest Disease Detection and Binary Mask Generation

The infectious disease tuberculosis (TB) continues to pose a serious risk to global health specially in developing world. There are over 10 million new cases of tuberculosis each year. Machine and deep learning models are trained to recognize specific pixels inside a medical image for the purposes of classification and disease progression tracking, however the decision-making mechanism of these models is hidden from the user. In this context, Explainable artificial intelligence (XAI) refers to strategies that allow humans to understand the results of AI algorithms. Recently, a variety of XAI methods for classification and generative have been proposed; however, these methods only use a subset of most discriminative power characteristics, resulting in false positives. This article proposes CycleGAN-based multi-functional generative adversarial networks to efficiently solve these challenges. Proposed model is trained in weakly supervised context to identify and visualize the disease effects and finally generate binary mask in data augmentation context. The model takes a Chest radiography (CXR) as input, creates a change map showing the disease's effect at a specific spot, and then uses this map to create a binary mask of original image. Results on publicly available TB dataset, TBX11K, confirm that the proposed model produces highly accurate result.