Chest X-rays are a common diagnostic tool, but overlapping ribs and collarbones can hide signs of disease. The clinical standard for removing these bone shadows, dual-energy subtraction, requires special equipment and extra radiation, limiting its use. Generative models have been explored as an alternative, but they often struggle with image quality and need large amounts of paired training data.
The proposed DeBoneDiT architecture tackles these issues in two ways. First, a depth auto-encoder compresses the image into a latent space while using depth features from a pretrained model to keep anatomical structure intact. Second, the method treats bone suppression as a Brownian bridge diffusion process, where the original X-ray serves as both the starting point and a structural guide, narrowing the gap between source and target domains.
To support training, the team built and released SZCH-X-Rays, which they say is the largest public paired dataset for this task, containing 741 chest X-ray and soft-tissue image pairs. In experiments, DeBoneDiT achieved better bone suppression than prior approaches while using less computation, according to the paper.