Although radiographs are the most frequently used medical imaging modality worldwide due to their cost-effectiveness and widespread accessibility, the structural superposition along the x-ray paths often renders lung nodules difficult to detect. In this study, we apply “X-ray dissectography” to dissect the lungs digitally in a few radiographic projections, suppress the interference of irrelevant structures, and improve lung nodule detectability. Then, we design a novel collaborative detection network to localize lung nodules in both the dissected 2D projections and the 3D physical space. Our experimental results show that our approach can significantly improve the average precision by 20+% in comparison with detecting lung nodules from the original projections using a popular detection network. Our proposed approach and results suggest a potential in re-designing the current X-ray imaging protocols and workflows and improving the diagnostic performance of chest radiographs in lung diseases.
|