GAN-based bone suppression using a combined loss function

Accurate analysis of chest radiographs (X-rays) is essential for diagnosing diseases such as pneumonia and lung cancer, yet bone structures often obscure critical soft tissues and lesions. From an artificial intelligence perspective, bone suppression can be formulated using different modeling paradigms that reflect distinct assumptions about the task.

In this study, the problem is addressed as a comparative methodological investigation, and three conceptually different approaches are systematically evaluated within a unified experimental framework: denoising-based regression using autoencoders, structured image-to-image transformation using U-Net architectures, and distribution-based generative modeling using adversarial learning. In addition, the impact of different loss configurations and training regimes on reconstruction quality is examined. An enhanced generative adversarial network (GAN) with improved generator and discriminator components and a combined loss function (Wasserstein, L1, perceptual, and Sobel losses) is proposed to improve structural consistency and preserve soft-tissue appearance.

Model performance was assessed using the peak signal-to-noise ratio (PSNR) and the multi-scale structural similarity index measure (MS-SSIM). Among the evaluated approaches, the GAN achieved the best performance, reaching a PSNR of 44.09 dB and an MS-SSIM of 0.9968, and outperformed recently published methods evaluated on the same dataset.

These results highlight the importance of both modeling paradigm selection and loss formulation for achieving structurally consistent bone suppression in chest radiographs.