Investigating the impact of novel XRayGAN in feature extraction for thoracic disease detection in chest radiographs: lung cancer

Lung cancer remains one of the most lethal malignancies worldwide, underscoring the urgent need for early detection and intervention to improve survival rates. While computed tomography (CT) has been crucial in identifying and diagnosing lung diseases, concerns surrounding radiation exposure and expenses have drawn attention to alternative diagnostic methods. Addressing this need, our research explores the potential of X-ray imaging combined with deep learning algorithms for early lung cancer diagnosis. Developing a reliable deep learning (DL) model hinges upon the availability of an adequate and high-quality dataset. However, the challenges in obtaining such datasets in medical imaging, including privacy concerns, access restrictions, and data labeling requirements, have been substantial. To overcome these hurdles, we have created the XRayGAN, a synthetic X-ray image generator, which offers a solution for generating diverse and high-quality X-ray images. The study investigates features extracted from chest radiographs using the novel XRayGAN to improve thoracic disease detection. XRayGAN leverages a variant of the auxiliary classifier generative adversarial network (ACGAN) and introduces a novel approach using extracted image features from a separate multiscale feature learning module as input labels for both the generator and discriminator. Additionally, our generative adversarial network (GAN) employs novel loss functions, maintaining constant weights through gradient adjustment to ensure model stability. Our study utilizes two types of datasets as inputs—the publicly available NIH dataset and a self-collected dataset. We have observed promising results from the proposed XRayGAN, particularly in image generation quality and improved accuracy in thoracic disease classification. Evaluation metrics such as accuracy, precision, and recall have been utilized to measure and validate the model’s performance. Furthermore, to ensure the practicality of our approach, a self-turing test was performed with the aid of an expert radiologist.