Fig. 3From: Deep learning tomographic reconstruction through hierarchical decomposition of domain transformsIllustration of the concept of hierarchical CT reconstruction. The diagrams in the outer ring represent the input data, i.e., line integrals at various rotation angles and radial offsets. The diagrams in the middle ring represent data in the intermediate domain, i.e., partial line integrals. The diagram at the center represents the output, i.e., the reconstructed image. The gray connections illustrate the flow of data through the network. As data go through the hierarchy, the depth resolution improves, while the number of angles decreases. The final reconstructed image is formed when the number of depth bins equals the desired size of the reconstructed image, and the number of angles reaches unity. The key benefit of hierarchical reconstruction is that the elementary reconstruction steps in each hierarchical stage are relatively localized, making it suitable for efficient implementation as neural networks. The red rectangles illustrate the localized correspondence across hierarchical stagesBack to article page