Abstract
Recent advancements in deep learning have led to significant improvements in pneumoconiosis diagnosis from chest X-rays (CXR). However, these models typically require large training datasets, which are challenging to collect due to the rarity of the disease and strict data-sharing limitations. In addition, the process of annotating medical images is labor-intensive, requiring highly skilled radiologists, which further limits data availability. To address this data scarcity, we propose a novel approach to generate synthetic pathology in CXR images, thus augmenting existing datasets and improving model training efficiency. While bidirectional generative adversarial networks (GAN), such as CycleGAN, can perform image translation between domains without paired samples, these methods often struggle to maintain structural and pathological consistency, especially with fine details. This study introduces a multiscale attention-based GAN (MSA-GAN) model that enhances CycleGAN with a multiscale attention generator, local-global discriminators and structural similarity (SSIM) loss, ensuring greater fidelity in preserving structural and pathological details during translation. We utilise MSA-GAN-generated synthetic CXR images to train CNN models for pneumoconiosis classification and lung segmentation tasks. Experimental results indicate that CNN trained on MSA-GAN-generated images outperforms existing CNN-based methods, showing improved accuracy and consistency in both qualitative and quantitative assessments across classification and segmentation tasks.